Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,6 @@
+__pycache__/
+*.py[cod]
+.DS_Store
+*.egg-info/
+dist/
+build/

LICENSE

@@ -0,0 +1,21 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   Copyright 2026 MAGI Labs
+
+   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.
+
+---
+Attribution Notices:
+- Pretraining data includes leakage-audited series from Salesforce/GiftEvalPretrain (Apache-2.0).

README.md

@@ -0,0 +1,79 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: time-series-forecasting
+tags:
+  - time-series
+  - time-series-forecasting
+  - probabilistic-forecasting
+  - jepa
+  - world-model
+  - foundation-model
+  - pytorch
+  - safetensors
+datasets:
+  - Salesforce/GiftEvalPretrain
+---
+
+# Telos-Small (`50.1M`) — Time-Series JEPA Foundation Model
+
+**Telos-Small** (`50,100,747` parameters) is a standalone zero-shot **Joint-Embedding Predictive Architecture (JEPA)** for probabilistic time-series forecasting developed by **MAGI Labs**.
+
+Instead of autoregressively generating raw time-series tokens step-by-step, **Telos** encodes observed history into patch representations, predicts future trajectory representations directly in latent space across up to **32 horizon patches (`1,024` steps) in a single forward pass**, and projects those predicted latents onto **21 calibrated quantiles (`0.01 .. 0.99`)**.
+
+---
+
+## How Telos Works
+
+1. **Causal Welford + `asinh` Patch Scaling (`patch_size = 32`):** Each 32-step patch is normalized using running causal Welford statistics and inverse-hyperbolic-sine (`asinh`) compression, handling heavy-tailed spikes, zero-inflated series, and arbitrary scale shifts without lookahead leakage.
+2. **Two-Stream Encoder with U-Net Skips & Value-Residual Attention (`12` blocks, `d_model = 512`):** Alternates three causal time-attention blocks with one cross-variate block (`[time, time, time, time_variate] × 3`) using `SinCosRoPE`, sigmoid-gated SDPA, U-Net skip connections (`layer[k] -> layer[11-k]`), and first-layer Value-Residual links across contexts up to **`4,096`–`8,192` steps**.
+3. **Horizon-Conditioned JEPA Latent Predictor (`2` blocks):** Given causal context latents and horizon step embeddings, predicts future target-encoder patch latents across **`1,024` future steps in one shot**, trained with a momentum EMA target encoder (`V-JEPA`), Epps-Pulley isotropic Gaussian regularization (`SIGReg`), multi-origin **Tiled Decoupled-Horizon Forecasting (`Tiled-DHF`)**, and orthogonalized **`NorMuon` (`Newton-Schulz-5`) + `AdamW`**.
+4. **Monotone 21-Knot Quantile Readout + Advancing Welford Frames:** Outputs strictly non-crossing quantiles `(0.01, 0.05, 0.10, ..., 0.90, 0.95, 0.99)` supporting univariate series, multivariate panels, past-only covariates, and known-future covariates.
+
+---
+
+## Quickstart (`transformers`)
+
+```bash
+pip install torch numpy safetensors transformers
+```
+
+```python
+import numpy as np
+from transformers import AutoModel
+
+model = AutoModel.from_pretrained("magilabsco/Telos-Small", trust_remote_code=True)
+
+# 1D series, (B, T) batch, or (B, V, T) multivariate panel -> returns (B, V, H, 21) quantiles
+context = np.sin(np.linspace(0, 16 * np.pi, 512, dtype=np.float32)) + 100.0
+quantiles = model.forecast(context, prediction_length=96)  # shape: (1, 1, 96, 21)
+median_forecast = quantiles[0, 0, :, 10]                   # 0.50 quantile
+```
+
+---
+
+## Zero-Shot Benchmark Results (`402` Tasks Across 4 Suites)
+
+Evaluated zero-shot across **402 tasks** on four public benchmarks (higher Skill % vs. Seasonal Naive is better):
+
+| Benchmark Suite | Tasks / Scale | Metric | **Telos-Small (`50.1M` Pure JEPA)** | **Telos-Small (`50.1M` + Analytical Blend)** | Notable Reference Baselines |
+| :--- | :--- | :--- | :---: | :---: | :--- |
+| **Salesforce `GIFT-Eval` (Full Panel)** | `97` tasks (`371,330` windows) | CRPS Skill % | **`43.21%`** | **`43.79%`** | Beats `PatchTST` (`41.30%`), `TFT` (`39.53%`), `DLinear` (`15.44%`) |
+| **AutoGluon `FEV-Bench` (`fev-bench-mini`)** | `20` multi-domain & covariate tasks | SQL Skill % | **`39.49%`** | **`42.82%`** | Beats `TFT` (`35.26%`), `PatchTST` (`34.96%`), `CatBoost` (`28.41%`) |
+| **Cognite `IND-Eval` (Industrial)** | `187` sensor & well tasks (`9` sources) | CRPS Skill % | **`38.18%`** | `34.06%` | Pure JEPA beats analytical blend by `+4.12` pts |
+| **ICML-2026 `TIME` Benchmark** | `98` zero-overlap fresh tasks (`50` datasets) | CRPS Skill % | **`36.54%`** | `35.70%` | Beats `Sundial-base` (`128M`, `33.66%`); trails `Chronos-Bolt` (`205M`, `38.00%`) by `1.46` pts |
+
+---
+
+## Model Specifications
+
+- **Model Name:** `Telos-Small`
+- **Parameters:** `50,100,747` (`191 MB` in `fp32` `model.safetensors`)
+- **Architecture:** `12` Two-Stream Transformer Encoder blocks (`d_model=512`, `8` heads, `mlp_hidden_dim=1152`, U-Net skips, Value-Residual) + `2` `JEPALatentPredictor` blocks (`hidden=1024`) + `QuantileHead` (`21` knots)
+- **Patch Size / Single-Shot Horizon:** `32` steps per patch / up to `32` patches (`1,024` steps) per single forward shot
+- **Context Length:** `4,096` steps default (`8,192` max)
+- **Training Compute:** `43,591` DDP steps (`12,570,406` forecast origins) on `2× NVIDIA Tesla P100-16GB` GPUs
+
+## License
+
+Released under the [Apache-2.0 License](LICENSE). Pretraining corpus incorporates audited series from `Salesforce/GiftEvalPretrain` (Apache-2.0) and synthetic physical dynamical priors (Copyright 2026 MAGI Labs).

__init__.py

@@ -0,0 +1,24 @@
+"""Telos (`Telos-Small`, 50.1M Parameters): Standalone Time-Series JEPA Foundation Model."""
+
+from .configuration_telos import QUANTILE_KNOTS_21, TelosConfig
+from .modeling_telos import (
+    ROLE_KNOWN_COV,
+    ROLE_PAST_COV,
+    ROLE_TARGET,
+    TelosForForecasting,
+    TelosModel,
+    load_telos_checkpoint,
+    save_telos_checkpoint,
+)
+
+__all__ = [
+    "QUANTILE_KNOTS_21",
+    "ROLE_KNOWN_COV",
+    "ROLE_PAST_COV",
+    "ROLE_TARGET",
+    "TelosConfig",
+    "TelosForForecasting",
+    "TelosModel",
+    "load_telos_checkpoint",
+    "save_telos_checkpoint",
+]

config.json

@@ -0,0 +1,78 @@
+{
+  "attend_missing_patches": true,
+  "attn_bias": false,
+  "block_schedule": [
+    "time",
+    "time",
+    "time",
+    "time_variate",
+    "time",
+    "time",
+    "time",
+    "time_variate",
+    "time",
+    "time",
+    "time",
+    "time_variate"
+  ],
+  "block_size": 32,
+  "context_length": 4096,
+  "covariate_enabled": true,
+  "d_model": 512,
+  "head_hidden_dim": 1792,
+  "jepa_predictor_hidden": 1024,
+  "jepa_predictor_layers": 2,
+  "max_context_length": 8192,
+  "minimum_scale": 1e-06,
+  "mlp_bias": false,
+  "mlp_hidden_dim": 1152,
+  "model_name": "Telos-Small",
+  "model_type": "telos",
+  "norm_eps": 1e-06,
+  "num_heads": 8,
+  "patch_proj_hidden": 1536,
+  "patch_size": 32,
+  "per_shot_horizon_patches": 32,
+  "quantile_knots": [
+    0.01,
+    0.05,
+    0.1,
+    0.15,
+    0.2,
+    0.25,
+    0.3,
+    0.35,
+    0.4,
+    0.45,
+    0.5,
+    0.55,
+    0.6,
+    0.65,
+    0.7,
+    0.75,
+    0.8,
+    0.85,
+    0.9,
+    0.95,
+    0.99
+  ],
+  "quantile_real_cap_k": 6.0,
+  "quantile_width_factor": 3.0,
+  "rope_partial_factor": 0.5,
+  "rope_theta": 10000.0,
+  "split_context_horizon_attn": true,
+  "stabilize_with_global": false,
+  "time_attn_gate": "sdpa_sigmoid",
+  "transformers_version": "5.22.2",
+  "unet_skip_connections": true,
+  "use_asinh": true,
+  "value_residual": true,
+  "variate_attn_gate": "sdpa_sigmoid",
+  "architectures": [
+    "TelosModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_telos.TelosConfig",
+    "AutoModel": "modeling_telos.TelosModel"
+  }
+}

configuration_telos.py

@@ -0,0 +1,110 @@
+"""Hugging Face `PreTrainedConfig` for Telos (`Telos-Small`, 50.1M parameters)."""
+
+from __future__ import annotations
+
+from typing import List, Sequence, Tuple
+
+from transformers import PretrainedConfig
+
+QUANTILE_KNOTS_21: Tuple[float, ...] = (
+    0.01, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45,
+    0.50,
+    0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95, 0.99,
+)
+
+
+class TelosConfig(PretrainedConfig):
+    """Configuration class for `TelosModel` (`Telos-Small`, 50.1M Time-Series JEPA Foundation Model)."""
+
+    model_type = "telos"
+
+    def __init__(
+        self,
+        model_name: str = "Telos-Small",
+        patch_size: int = 32,
+        d_model: int = 512,
+        num_heads: int = 8,
+        mlp_hidden_dim: int = 1152,
+        patch_proj_hidden: int = 1536,
+        head_hidden_dim: int = 1792,
+        block_schedule: Sequence[str] = (
+            "time", "time", "time", "time_variate",
+            "time", "time", "time", "time_variate",
+            "time", "time", "time", "time_variate",
+        ),
+        quantile_knots: Sequence[float] = QUANTILE_KNOTS_21,
+        block_size: int = 32,
+        context_length: int = 4096,
+        max_context_length: int = 8192,
+        rope_theta: float = 10000.0,
+        rope_partial_factor: float = 0.5,
+        norm_eps: float = 1e-6,
+        attn_bias: bool = False,
+        mlp_bias: bool = False,
+        time_attn_gate: str = "sdpa_sigmoid",
+        variate_attn_gate: str = "sdpa_sigmoid",
+        minimum_scale: float = 1e-6,
+        use_asinh: bool = True,
+        stabilize_with_global: bool = False,
+        covariate_enabled: bool = True,
+        split_context_horizon_attn: bool = True,
+        attend_missing_patches: bool = True,
+        quantile_width_factor: float = 3.0,
+        per_shot_horizon_patches: int = 32,
+        quantile_real_cap_k: float = 6.0,
+        jepa_predictor_layers: int = 2,
+        jepa_predictor_hidden: int = 1024,
+        unet_skip_connections: bool = True,
+        value_residual: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.model_name = model_name
+        self.patch_size = int(patch_size)
+        self.d_model = int(d_model)
+        self.num_heads = int(num_heads)
+        self.mlp_hidden_dim = int(mlp_hidden_dim)
+        self.patch_proj_hidden = int(patch_proj_hidden)
+        self.head_hidden_dim = int(head_hidden_dim)
+        self.block_schedule: List[str] = list(block_schedule)
+        self.quantile_knots: List[float] = [float(q) for q in quantile_knots]
+        self.block_size = int(block_size)
+        self.context_length = int(context_length)
+        self.max_context_length = int(max_context_length)
+        self.rope_theta = float(rope_theta)
+        self.rope_partial_factor = float(rope_partial_factor)
+        self.norm_eps = float(norm_eps)
+        self.attn_bias = bool(attn_bias)
+        self.mlp_bias = bool(mlp_bias)
+        self.time_attn_gate = str(time_attn_gate)
+        self.variate_attn_gate = str(variate_attn_gate)
+        self.minimum_scale = float(minimum_scale)
+        self.use_asinh = bool(use_asinh)
+        self.stabilize_with_global = bool(stabilize_with_global)
+        self.covariate_enabled = bool(covariate_enabled)
+        self.split_context_horizon_attn = bool(split_context_horizon_attn)
+        self.attend_missing_patches = bool(attend_missing_patches)
+        self.quantile_width_factor = float(quantile_width_factor)
+        self.per_shot_horizon_patches = int(per_shot_horizon_patches)
+        self.quantile_real_cap_k = float(quantile_real_cap_k)
+        self.jepa_predictor_layers = int(jepa_predictor_layers)
+        self.jepa_predictor_hidden = int(jepa_predictor_hidden)
+        self.unet_skip_connections = bool(unet_skip_connections)
+        self.value_residual = bool(value_residual)
+
+    @property
+    def num_quantiles(self) -> int:
+        return len(self.quantile_knots)
+
+    @property
+    def median_index(self) -> int:
+        return list(self.quantile_knots).index(0.5)
+
+    @property
+    def head_dim(self) -> int:
+        return self.d_model // self.num_heads
+
+    @property
+    def rope_dim(self) -> int:
+        rd = int(self.head_dim * self.rope_partial_factor)
+        return max(2, rd - (rd % 2))

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3a5434cf774848efab77b1ef4f8189358694f1374fd715b31f7ce88eeee7305
+size 200417852

modeling_telos.py

@@ -0,0 +1,1432 @@
+#!/usr/bin/env python3
+"""Telos (`Telos-Small`, 50.1M Parameters): Standalone Time-Series Joint-Embedding Predictive Architecture (JEPA).
+
+Architecture highlights:
+  1. Causal per-patch Welford mean/std normalization with `asinh` compression and
+     `advancing_block_frames` multi-patch horizon refinement.
+  2. 12-block Two-Stream `[Enc ; Dec]` Transformer (`d_model=512`, `num_heads=8`,
+     `[time, time, time, time_variate] x 3`) with `SinCosRoPE`, `sdpa_sigmoid` gated
+     time/variate attention, U-Net skip connections, Value-Residual attention, and `FusedSwiGLU`.
+  3. 2-block Horizon-Conditioned `JEPALatentPredictor` predicting future patch representations
+     in latent space across up to `32` patches (`1,024` steps) in a single forward shot.
+  4. 21-knot monotone `QuantileHead` (`0.01 .. 0.99`) with full support for univariate,
+     multivariate, past covariates, and known-future covariates.
+"""
+
+from __future__ import annotations
+
+import functools
+import math
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from transformers import PreTrainedModel
+
+try:
+    from .configuration_telos import QUANTILE_KNOTS_21, TelosConfig
+except ImportError:
+    from configuration_telos import QUANTILE_KNOTS_21, TelosConfig  # type: ignore
+
+ROLE_TARGET = 0
+ROLE_PAST_COV = 1
+ROLE_KNOWN_COV = 2
+
+LAYER_KIND_AXES: Dict[str, Tuple[bool, ...]] = {
+    "time": (True,),
+    "variate": (False,),
+    "time_variate": (True, False),
+}
+
+
+# ==============================================================================
+# 1. Causal Welford & Asinh Scaler Primitives
+# ==============================================================================
+
+
+def _fp32_io(*dtypes: torch.dtype) -> Tuple[torch.dtype, torch.dtype]:
+    out_dtype = functools.reduce(torch.promote_types, dtypes)
+    return out_dtype, torch.promote_types(out_dtype, torch.float32)
+
+
+def compute_causal_statistics(
+    data: Tensor,
+    observation_mask: Tensor,
+    minimum_scale: float,
+    stabilize_with_global: bool,
+    scale_factor_exponent: float = 10.0,
+) -> Tuple[Tensor, Tensor]:
+    weights = torch.ones_like(data) * observation_mask
+    with torch.no_grad():
+        x = data.to(torch.float64)
+        w = weights.to(torch.float64)
+        weighted = w * x
+        cum_w = torch.cumsum(w, dim=-1)
+        cum_x = torch.cumsum(weighted, dim=-1)
+        denominator = cum_w.clamp_min(1.0)
+        means = cum_x / denominator
+        shifted = torch.zeros_like(means)
+        shifted[..., 1:] = means[..., :-1]
+        increment = (x - shifted) * (x - means) * w
+        m2 = torch.cumsum(increment, dim=-1)
+        variance = m2 / torch.clamp(denominator - 1.0, min=1.0)
+        scales = torch.sqrt(variance).clamp(min=minimum_scale)
+        if stabilize_with_global:
+            lo, hi = 10.0 ** (-scale_factor_exponent), 10.0 ** scale_factor_exponent
+            global_den = (weights * observation_mask).sum(-1, keepdim=True).clamp_min(1.0)
+            global_means = torch.nan_to_num(weighted.sum(-1, keepdim=True) / global_den)
+            global_var = (((x - global_means) * weights * observation_mask) ** 2).sum(-1, keepdim=True) / global_den
+            global_scale = torch.sqrt(global_var + minimum_scale).expand_as(scales)
+            floor = torch.max(torch.tensor(minimum_scale, device=scales.device, dtype=scales.dtype), global_scale * lo)
+            scales = torch.clamp(scales, min=floor, max=global_scale * hi)
+        return means.to(data.dtype), scales.to(data.dtype)
+
+
+def causal_patch_loc_scale(
+    data: Tensor,
+    observation_mask: Tensor,
+    patch_size: int,
+    minimum_scale: float,
+    stabilize_with_global: bool,
+) -> Tuple[Tensor, Tensor]:
+    with torch.no_grad():
+        means, scales = compute_causal_statistics(data, observation_mask, minimum_scale, stabilize_with_global)
+        patch_means = means.unfold(-1, patch_size, patch_size)[..., -1]
+        patch_scales = scales.unfold(-1, patch_size, patch_size)[..., -1]
+        return patch_means.repeat_interleave(patch_size, dim=-1), patch_scales.repeat_interleave(patch_size, dim=-1)
+
+
+def apply_affine_asinh_masked(
+    data: Tensor,
+    loc: Tensor,
+    scale: Tensor,
+    observation_mask: Tensor,
+    use_asinh: bool,
+) -> Tensor:
+    out_dtype, work = _fp32_io(data.dtype, loc.dtype, scale.dtype)
+    scaled = (data.to(work) - loc.to(work)) / scale.to(work).clamp_min(torch.finfo(work).eps)
+    scaled = scaled.masked_fill(~observation_mask, 0.0)
+    if use_asinh:
+        scaled = torch.where(observation_mask, scaled.asinh(), scaled)
+    return scaled.to(out_dtype)
+
+
+def invert_affine_asinh(x: Tensor, loc: Tensor, scale: Tensor, use_asinh: bool) -> Tensor:
+    out_dtype, work = _fp32_io(x.dtype, loc.dtype, scale.dtype)
+    x = x.to(work)
+    if use_asinh:
+        x = x.clamp(-6.0, 6.0).sinh()
+    return (x * scale.to(work) + loc.to(work)).to(out_dtype)
+
+
+def clamp_nonfinite(vals: Tensor) -> Tensor:
+    clamped = torch.where(
+        vals == float("inf"),
+        torch.where(vals.isfinite(), vals, -float("inf")).amax(dim=-1, keepdim=True),
+        torch.where(
+            vals == -float("inf"),
+            torch.where(vals.isfinite(), vals, float("inf")).amin(dim=-1, keepdim=True),
+            vals,
+        ),
+    )
+    return torch.where(clamped.isinf(), torch.zeros_like(clamped), clamped)
+
+
+def denormalize_quantiles(preds: Tensor, loc: Tensor, scale: Tensor, use_asinh: bool) -> Tensor:
+    return clamp_nonfinite(invert_affine_asinh(preds, loc, scale, use_asinh))
+
+
+def welford_combine(
+    w_a: Tensor, mean_a: Tensor, m2_a: Tensor, w_b: Tensor, mean_b: Tensor, m2_b: Tensor
+) -> Tuple[Tensor, Tensor, Tensor]:
+    w = w_a + w_b
+    w_safe = w.clamp(min=torch.finfo(w.dtype).tiny)
+    delta = mean_b - mean_a
+    mean = mean_a + delta * w_b / w_safe
+    m2 = m2_a + m2_b + delta.square() * w_a * w_b / w_safe
+    is_zero = w == 0
+    mean = torch.where(is_zero, torch.zeros_like(mean), mean)
+    m2 = torch.where(is_zero, torch.zeros_like(m2), m2)
+    return w, mean, m2
+
+
+def cumulative_welford(w: Tensor, mean: Tensor, m2: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
+    n = w.shape[1]
+    d = 1
+    while d < n:
+        pad = (d, 0)
+        w_l = F.pad(w[:, : n - d], pad)
+        mean_l = F.pad(mean[:, : n - d], pad)
+        m2_l = F.pad(m2[:, : n - d], pad)
+        w, mean, m2 = welford_combine(w_l, mean_l, m2_l, w, mean, m2)
+        d *= 2
+    return w, mean, m2
+
+
+def advancing_block_frames(
+    context: Tensor, observation_mask: Tensor, sinh_median: Tensor, minimum_scale: float
+) -> Tuple[Tensor, Tensor]:
+    r, n_pred, ps = sinh_median.shape
+    blank = sinh_median.new_zeros(r, n_pred, ps)
+    values = torch.cat([context.unflatten(-1, (-1, ps)).to(torch.float64), blank], dim=1)
+    weights = torch.cat([observation_mask.unflatten(-1, (-1, ps)).to(torch.float64), blank], dim=1)
+    cnt = weights.sum(-1)
+    mean = (values * weights).sum(-1) / cnt.clamp(min=1.0)
+    m2 = (weights * (values - mean.unsqueeze(-1)).square()).sum(-1)
+    is_zero = cnt == 0
+    mean = torch.where(is_zero, torch.zeros_like(mean), mean)
+    m2 = torch.where(is_zero, torch.zeros_like(m2), m2)
+    cnt, mean, m2 = (s[:, -n_pred - 1] for s in cumulative_welford(cnt, mean, m2))
+
+    offset = mean
+    clip = torch.finfo(torch.float32).max ** 0.5 / 1e3
+    locs, scales = [], []
+    for h in range(n_pred):
+        var = (m2 / (cnt - 1.0).clamp(min=1.0)).clamp(min=0.0)
+        scale = var.sqrt().clamp(min=minimum_scale)
+        locs.append(mean)
+        scales.append(scale)
+        folded = (sinh_median[:, h] * scale[:, None] + mean[:, None] - offset[:, None]).clamp(-clip, clip)
+        folded = folded + offset[:, None]
+        b_mean = folded.mean(-1)
+        b_m2 = (folded - b_mean[:, None]).square().sum(-1)
+        cnt, mean, m2 = welford_combine(cnt, mean, m2, torch.full_like(cnt, ps), b_mean, b_m2)
+    return torch.stack(locs, dim=1), torch.stack(scales, dim=1)
+
+
+def instance_norm_stats(values: Tensor, observation_mask: Tensor, minimum_scale: float) -> Tuple[Tensor, Tensor, Tensor]:
+    orig_dtype = values.dtype
+    stat_dtype = torch.promote_types(orig_dtype, torch.float32)
+    x = values.to(stat_dtype)
+    mask_f = observation_mask.to(stat_dtype)
+    n = mask_f.sum(dim=-1, keepdim=True).clamp(min=1.0)
+    loc = (x * mask_f).sum(dim=-1, keepdim=True) / n
+    var = (((x - loc) * mask_f) ** 2).sum(dim=-1, keepdim=True) / n
+    scale = var.sqrt().clamp_min(minimum_scale)
+    is_zero_one = ((values == 0.0) | (values == 1.0) | ~observation_mask).all(dim=-1, keepdim=True)
+    is_binary = is_zero_one & observation_mask.any(dim=-1, keepdim=True)
+    loc = torch.where(is_binary, torch.zeros_like(loc), loc)
+    scale = torch.where(is_binary, torch.ones_like(scale), scale)
+    return loc.to(orig_dtype), scale.to(orig_dtype), is_binary
+
+
+def ffill_with_mask(x: Tensor, mask: Tensor) -> Tensor:
+    idx = torch.arange(x.shape[-1], device=x.device).view(*([1] * (x.dim() - 1)), -1)
+    valid_idx = torch.where(mask, idx, torch.zeros_like(idx)).cummax(dim=-1).values
+    filled = torch.gather(x, -1, valid_idx)
+    return torch.where(mask.cummax(dim=-1).values.bool(), filled, torch.zeros_like(x))
+
+
+def pad_to_patch(rows: Tensor, mask: Tensor, patch_size: int) -> Tuple[Tensor, Tensor]:
+    rem = rows.shape[-1] % patch_size
+    if rem == 0:
+        return rows, mask
+    pad_len = patch_size - rem
+    return F.pad(rows, (pad_len, 0), value=0.0), F.pad(mask, (pad_len, 0), value=False)
+
+
+# ==============================================================================
+# 2. Two-Stream Masking & RoPE Gated Transformer Layers
+# ==============================================================================
+
+
+@dataclass(frozen=True)
+class TwoStreamMask:
+    mask: Tensor
+    enc_len: int
+
+
+def compute_enc_key_validity(obs_per_patch: Tensor, attend_missing: bool) -> Tensor:
+    if not attend_missing:
+        return obs_per_patch
+    started = obs_per_patch.cummax(dim=-1).values
+    return obs_per_patch | started
+
+
+def enc_causal_mask(key_valid: Tensor) -> Tensor:
+    _, ll = key_valid.shape
+    device = key_valid.device
+    causal = torch.tril(torch.ones(ll, ll, dtype=torch.bool, device=device))
+    eye = torch.eye(ll, dtype=torch.bool, device=device)
+    return (causal[None] & key_valid[:, None, :]) | eye[None]
+
+
+def build_forecast_time_mask(context_len: int, horizon: int, key_valid: Tensor, device: torch.device) -> Tensor:
+    rows = key_valid.shape[0]
+    enc = enc_causal_mask(key_valid)
+    enc_dec = torch.zeros(rows, context_len, horizon, dtype=torch.bool, device=device)
+    dec_enc = key_valid[:, None, :].expand(rows, horizon, context_len)
+    dec_dec = torch.tril(torch.ones(horizon, horizon, dtype=torch.bool, device=device))[None].expand(rows, horizon, horizon)
+    top = torch.cat([enc, enc_dec], dim=2)
+    bottom = torch.cat([dec_enc, dec_dec], dim=2)
+    return torch.cat([top, bottom], dim=1).unsqueeze(1)
+
+
+def build_forecast_positions(context_len: int, horizon: int, first_obs: Tensor) -> Tensor:
+    p = torch.arange(context_len + horizon, device=first_obs.device)
+    return p[None, :] - first_obs[:, None]
+
+
+def group_first_observed_patch(obs_per_patch: Tensor, inverse: Tensor, num_groups: int) -> Tensor:
+    n = obs_per_patch.shape[-1]
+    idx = inverse[:, None].expand(-1, n)
+    group_obs = torch.zeros(num_groups, n, dtype=torch.float32, device=obs_per_patch.device)
+    group_obs.scatter_reduce_(0, idx, obs_per_patch.float(), reduce="amax", include_self=True)
+    return group_obs.argmax(dim=-1)
+
+
+def build_two_stream_positions(n_patches: int, first_obs: Tensor) -> Tensor:
+    """`(R, 2N)` positions: encoder and decoder token of the same patch share a time."""
+    p = torch.arange(n_patches, device=first_obs.device)
+    row = p[None, :] - first_obs[:, None]
+    return torch.cat([row, row], dim=-1)
+
+
+def compute_block_layout(min_context: Tensor, n_patches: int, block_size: int) -> Tuple[Tensor, Tensor]:
+    """Per decoder position: the anchor of its block and whether it is active (`p >= s`)."""
+    device = min_context.device
+    p = torch.arange(n_patches, device=device)[None, :]
+    s = min_context[:, None]
+    active = p >= s
+    offset = (p - s).clamp(min=0)
+    anchor = s + block_size * (offset // block_size)
+    anchor = torch.where(active, anchor, torch.zeros_like(anchor))
+    return anchor, active
+
+
+def build_two_stream_time_mask(
+    min_context: Tensor,
+    n_patches: int,
+    block_size: int,
+    key_valid: Tensor,
+    bidirectional_rows: Optional[Tensor],
+) -> Tensor:
+    """`(R, 1, 2N, 2N)` time mask over a full `[Enc ; Dec]` window (the covariate path)."""
+    anchor, active = compute_block_layout(min_context, n_patches, block_size)
+    rows, n = active.shape
+    device = active.device
+    causal = torch.tril(torch.ones(n, n, dtype=torch.bool, device=device))
+    key_col = key_valid[:, None, :]
+    j = torch.arange(n, device=device).view(1, 1, n)
+    eye = torch.eye(n, dtype=torch.bool, device=device)
+    enc_enc = enc_causal_mask(key_valid)
+    if bidirectional_rows is not None:
+        enc_bi = key_valid[:, None, :].expand(rows, n, n) | eye[None]
+        enc_enc = torch.where(bidirectional_rows[:, None, None], enc_bi, enc_enc)
+    enc_dec = torch.zeros(rows, n, n, dtype=torch.bool, device=device)
+    dec_enc = active[:, :, None] & (j < anchor[:, :, None]) & key_col
+    same_block = anchor[:, :, None] == anchor[:, None, :]
+    dec_dec = same_block & causal[None] & active[:, :, None] & active[:, None, :]
+    dec_dec = dec_dec | eye[None]
+    top = torch.cat([enc_enc, enc_dec], dim=2)
+    bottom = torch.cat([dec_enc, dec_dec], dim=2)
+    return torch.cat([top, bottom], dim=1).unsqueeze(1)
+
+
+def variate_firewall(
+    r: int, device: torch.device, target_rows: Optional[Tensor], bidirectional_rows: Optional[Tensor]
+) -> Optional[Tensor]:
+    """`(R, R)` forbidden query-to-key row pairs, or None."""
+    if target_rows is None and bidirectional_rows is None:
+        return None
+    forbid = torch.zeros(r, r, dtype=torch.bool, device=device)
+    if target_rows is not None:
+        forbid = forbid | ((~target_rows[:, None]) & target_rows[None, :])
+    if bidirectional_rows is not None:
+        forbid = forbid | (bidirectional_rows[:, None] & (~bidirectional_rows)[None, :])
+    return forbid
+
+
+_IDENTITY_VARIATE_MASK = "identity"
+
+
+def build_two_stream_variate_mask(
+    group_ids: Tensor,
+    enc_key_valid: Tensor,
+    n_dec: int,
+    target_rows: Optional[Tensor] = None,
+    bidirectional_rows: Optional[Tensor] = None,
+) -> Optional[Tensor | str]:
+    r, _ = enc_key_valid.shape
+    if r == 1:
+        return None
+    device = enc_key_valid.device
+    if int(torch.unique(group_ids).numel()) == r:
+        # Every row is an independent univariate series -> variate attention mask is I_r
+        return _IDENTITY_VARIATE_MASK
+    row_eq = group_ids[:, None] == group_ids[None, :]
+    dec_valid = torch.ones(r, n_dec, dtype=torch.bool, device=device)
+    slot_valid = torch.cat([enc_key_valid, dec_valid], dim=1)
+    key_valid = slot_valid.permute(1, 0)
+    mask = row_eq[None, None] & key_valid[:, None, None, :]
+    forbid = variate_firewall(r, device, target_rows, bidirectional_rows)
+    if forbid is not None:
+        mask = mask & ~forbid[None, None]
+    eye = torch.eye(r, dtype=torch.bool, device=device)
+    return mask | eye[None, None]
+
+
+def target_encode_covariate(
+    target_context: np.ndarray,
+    past_codes: np.ndarray,
+    future_codes: Optional[np.ndarray] = None,
+    smooth: float = 1.0,
+) -> Tuple[np.ndarray, Optional[np.ndarray]]:
+    """Per-series smoothed mean-target encoding for categorical/discrete covariates."""
+    target = np.asarray(target_context, dtype=np.float64)
+    past_codes = np.asarray(past_codes, dtype=np.float64)
+    observed = np.isfinite(target)
+    target = np.where(observed, target, 0.0)
+    item_count = float(observed.sum())
+    item_mean = float(target.sum() / item_count) if item_count > 0 else 0.0
+
+    categories = np.unique(past_codes[np.isfinite(past_codes)])
+    nan_slot = len(categories)
+
+    def _slots(codes: np.ndarray, unseen: Optional[int] = None) -> np.ndarray:
+        idx = np.searchsorted(categories, np.nan_to_num(codes, nan=0.0))
+        idx = np.minimum(idx, max(len(categories) - 1, 0))
+        known = np.isfinite(codes) & (len(categories) > 0)
+        if len(categories) > 0:
+            known &= categories[idx] == codes
+        return np.where(np.isnan(codes), nan_slot, np.where(known, idx, unseen if unseen is not None else 0)).astype(np.intp)
+
+    past_slots = _slots(past_codes)
+    sums = np.bincount(past_slots, weights=target * observed, minlength=nan_slot + 1)
+    counts = np.bincount(past_slots, weights=observed.astype(np.float64), minlength=nan_slot + 1)
+    lookup = np.append((smooth * item_mean + sums) / (smooth + counts), item_mean)
+
+    encoded_past = lookup[past_slots].astype(np.float32)
+    if future_codes is None:
+        return encoded_past, None
+    future_slots = _slots(np.asarray(future_codes, dtype=np.float64), unseen=nan_slot + 1)
+    return encoded_past, lookup[future_slots].astype(np.float32)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.eps = eps
+        self.scale = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        xf = x.float()
+        out = xf * torch.rsqrt(xf.pow(2).mean(-1, keepdim=True) + self.eps)
+        return (out * self.scale.float()).to(self.scale.dtype)
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, input_dims: int, hidden_dims: int, output_dims: int, use_bias: bool) -> None:
+        super().__init__()
+        self.hidden_layer = nn.Linear(input_dims, hidden_dims, bias=use_bias)
+        self.output_layer = nn.Linear(hidden_dims, output_dims, bias=use_bias)
+        self.residual_layer = nn.Linear(input_dims, output_dims, bias=use_bias)
+        self.activation = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.output_layer(self.activation(self.hidden_layer(x))) + self.residual_layer(x)
+
+
+class FusedSwiGLU(nn.Module):
+    def __init__(self, embed_dim: int, hidden_dim: int, use_bias: bool = False) -> None:
+        super().__init__()
+        self.w12 = nn.Linear(embed_dim, 2 * hidden_dim, bias=use_bias)
+        self.w3 = nn.Linear(hidden_dim, embed_dim, bias=use_bias)
+
+    def forward(self, x: Tensor) -> Tensor:
+        gate, up = self.w12(x).chunk(2, dim=-1)
+        prod = (gate.float() * F.silu(up.float())).clamp(min=-32768.0, max=32768.0).to(self.w3.weight.dtype)
+        return self.w3(prod)
+
+
+class SinCosRoPE(nn.Module):
+    inv_freq: Tensor
+
+    def __init__(self, dim: int, base: float, rotary_dim: int) -> None:
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        self.rotary_dim = rotary_dim
+        self.register_buffer("inv_freq", self._build_inv_freq(None), persistent=False)
+
+    def _build_inv_freq(self, device: Optional[torch.device]) -> Tensor:
+        inv_freq = torch.zeros(self.dim // 2, dtype=torch.float32, device=device)
+        idx = torch.arange(0, self.rotary_dim, 2, dtype=torch.float32, device=device)
+        inv_freq[: self.rotary_dim // 2] = 1.0 / (self.base ** (idx / self.rotary_dim))
+        return inv_freq
+
+    @torch.no_grad()
+    def forward(self, x: Tensor, position_ids: Tensor) -> Tuple[Tensor, Tensor]:
+        if self.inv_freq.is_meta or self.inv_freq.device != x.device or float(self.inv_freq[0]) != 1.0:
+            self.inv_freq = self._build_inv_freq(x.device)
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        return emb.cos().to(dtype=x.dtype), emb.sin().to(dtype=x.dtype)
+
+
+def _rotate_half(x: Tensor) -> Tensor:
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def _apply_rotary(q: Tensor, k: Tensor, cos: Tensor, sin: Tensor) -> Tuple[Tensor, Tensor]:
+    cos = cos.to(dtype=q.dtype).unsqueeze(2)
+    sin = sin.to(dtype=q.dtype).unsqueeze(2)
+    return (q * cos) + (_rotate_half(q) * sin), (k * cos) + (_rotate_half(k) * sin)
+
+
+def _sdpa(query: Tensor, key: Tensor, value: Tensor, mask: Optional[Tensor], is_causal: bool = False) -> Tensor:
+    batch, seqlen, _, _ = query.shape
+    out = F.scaled_dot_product_attention(
+        query.transpose(1, 2),
+        key.transpose(1, 2),
+        value.transpose(1, 2),
+        attn_mask=mask,
+        is_causal=is_causal,
+        dropout_p=0.0,
+    )
+    return out.transpose(1, 2).contiguous().view(batch, seqlen, -1)
+
+
+class Attention(nn.Module):
+    def __init__(
+        self, embed_dim: int, num_heads: int, is_time: bool, use_bias: bool, gate: str, value_residual: bool = False
+    ) -> None:
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        self.is_time = is_time
+        self.wQKV = nn.Linear(embed_dim, 3 * embed_dim, bias=use_bias)
+        self.wO = nn.Linear(embed_dim, embed_dim, bias=use_bias)
+        self._gate = nn.Linear(embed_dim, embed_dim, bias=use_bias) if gate == "sdpa_sigmoid" else None
+        self.v_res_logit = nn.Parameter(torch.tensor(-2.0)) if (value_residual and is_time) else None
+
+    def _qkv(self, x: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
+        batch, seq, _ = x.shape
+        qkv = self.wQKV(x.contiguous()).reshape(batch, seq, 3, self.head_dim, self.num_heads)
+        q, k, v = qkv.permute(2, 0, 1, 4, 3).unbind(dim=0)
+        return q.contiguous(), k.contiguous(), v.contiguous()
+
+    def forward(
+        self,
+        x: Tensor,
+        mask: Optional[Tensor | TwoStreamMask | str],
+        cos: Tensor,
+        sin: Tensor,
+        v_first: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        rows, num_patches, _ = x.shape
+        inputs = x if self.is_time else x.permute(1, 0, 2)
+        v_time_out: Optional[Tensor] = None
+        if not self.is_time and isinstance(mask, str) and mask == _IDENTITY_VARIATE_MASK:
+            qkv = self.wQKV(inputs.contiguous()).reshape(num_patches, rows, 3, self.head_dim, self.num_heads)
+            v = qkv[:, :, 2].permute(0, 1, 3, 2).contiguous().view(num_patches, rows, self.embed_dim)
+            out = v
+        else:
+            q, k, v = self._qkv(inputs)
+            if self.is_time:
+                v_time_out = v
+                if self.v_res_logit is not None and v_first is not None:
+                    lam = torch.sigmoid(self.v_res_logit).to(v.dtype)
+                    v = (1.0 - lam) * v + lam * v_first.to(v.dtype)
+                q, k = _apply_rotary(q, k, cos, sin)
+            if isinstance(mask, TwoStreamMask):
+                n = mask.enc_len
+                out_enc = _sdpa(q[:, :n], k[:, :n], v[:, :n], mask.mask[:, :, :n, :n])
+                out_dec = _sdpa(q[:, n:], k, v, mask.mask[:, :, n:, :])
+                out = torch.cat([out_enc, out_dec], dim=1)
+            else:
+                out = _sdpa(q, k, v, mask if not isinstance(mask, str) else None, is_causal=self.is_time and mask is None)
+        if self._gate is not None:
+            out = out * self._gate(inputs).sigmoid()
+        if self.is_time:
+            out = out.reshape(rows, num_patches, self.embed_dim)
+        else:
+            out = out.reshape(num_patches, rows, self.embed_dim).permute(1, 0, 2)
+        return self.wO(out), v_time_out
+
+
+class TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        kind: str,
+        embed_dim: int,
+        num_heads: int,
+        mlp_hidden_dim: int,
+        norm_eps: float,
+        attn_bias: bool,
+        mlp_bias: bool,
+        time_attn_gate: str,
+        variate_attn_gate: str,
+        value_residual: bool = False,
+    ) -> None:
+        super().__init__()
+        axes = LAYER_KIND_AXES[kind]
+        self.attn_norms = nn.ModuleList([RMSNorm(embed_dim, norm_eps) for _ in axes])
+        self.attns = nn.ModuleList(
+            [
+                Attention(
+                    embed_dim, num_heads, is_time, attn_bias,
+                    time_attn_gate if is_time else variate_attn_gate,
+                    value_residual=value_residual,
+                )
+                for is_time in axes
+            ]
+        )
+        self.mlp_norm = RMSNorm(embed_dim, norm_eps)
+        self.mlp = FusedSwiGLU(embed_dim, mlp_hidden_dim, mlp_bias)
+
+    def forward(
+        self,
+        x: Tensor,
+        time_mask: Optional[Tensor | TwoStreamMask],
+        variate_mask: Optional[Tensor],
+        cos: Tensor,
+        sin: Tensor,
+        v_first: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        res_dtype = self.mlp_norm.scale.dtype
+        x = x.to(res_dtype)
+        v_time_first = v_first
+        for norm, attn in zip(self.attn_norms, self.attns):
+            out, v_t = attn(norm(x), time_mask if attn.is_time else variate_mask, cos, sin, v_first=v_time_first)
+            if attn.is_time and v_time_first is None:
+                v_time_first = v_t
+            x = x + out.to(res_dtype)
+        return x + self.mlp(self.mlp_norm(x)).to(res_dtype), v_time_first
+
+
+class Transformer(nn.Module):
+    def __init__(self, config: TelosConfig) -> None:
+        super().__init__()
+        c = config
+        n_layers = len(c.block_schedule)
+        self.layers = nn.ModuleList(
+            [
+                TransformerBlock(
+                    kind=kind,
+                    embed_dim=c.d_model,
+                    num_heads=c.num_heads,
+                    mlp_hidden_dim=c.mlp_hidden_dim,
+                    norm_eps=c.norm_eps,
+                    attn_bias=c.attn_bias,
+                    mlp_bias=c.mlp_bias,
+                    time_attn_gate=c.time_attn_gate,
+                    variate_attn_gate=c.variate_attn_gate,
+                    value_residual=c.value_residual and (i > 0),
+                )
+                for i, kind in enumerate(c.block_schedule)
+            ]
+        )
+        self.unet_skip_scales: Optional[nn.Parameter] = (
+            nn.Parameter(torch.zeros(n_layers // 2, c.d_model)) if c.unet_skip_connections else None
+        )
+        self.rope = SinCosRoPE(dim=c.head_dim, base=c.rope_theta, rotary_dim=c.rope_dim)
+        self.out_norm = RMSNorm(c.d_model, c.norm_eps)
+
+    def forward(
+        self,
+        x: Tensor,
+        time_mask: Optional[Tensor | TwoStreamMask],
+        variate_mask: Optional[Tensor],
+        positions: Tensor,
+        return_layer_states: bool = False,
+    ) -> Tensor | Tuple[Tensor, List[Tensor]]:
+        cos, sin = self.rope(x, positions)
+        layer_states: List[Tensor] = []
+        early_skips: List[Tensor] = []
+        v_first: Optional[Tensor] = None
+        n_l = len(self.layers)
+        half = n_l // 2
+        for idx, layer in enumerate(self.layers):
+            if self.unet_skip_scales is not None and idx >= n_l - half:
+                skip_idx = n_l - 1 - idx
+                x = x + early_skips[skip_idx] * self.unet_skip_scales[skip_idx].to(x.dtype)
+            x, v_first = layer(x, time_mask, variate_mask, cos, sin, v_first=v_first)
+            if self.unet_skip_scales is not None and idx < half:
+                early_skips.append(x)
+            if return_layer_states:
+                layer_states.append(x)
+        out = self.out_norm(x)
+        if return_layer_states:
+            return out, layer_states
+        return out
+
+
+# ==============================================================================
+# 3. Patch Embedding & Monotone Quantile Head
+# ==============================================================================
+
+
+class PatchRepresentation(nn.Module):
+    def __init__(self, patch_size: int, d_model: int, hidden_dims: int) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_proj = ResidualBlock(2 * patch_size, hidden_dims, d_model, use_bias=True)
+        self.forecast_query = nn.Parameter(torch.zeros(d_model))
+
+    def forward(self, scaled: Tensor, observed: Tensor) -> Tensor:
+        r, t = scaled.shape
+        n = t // self.patch_size
+        values = scaled.reshape(r, n, self.patch_size)
+        missing = (~observed).reshape(r, n, self.patch_size).to(scaled.dtype)
+        x = torch.cat([values, missing], dim=-1).to(self.patch_proj.hidden_layer.weight.dtype)
+        return self.patch_proj(x)
+
+    def forecast_tokens(self, rows: int, n_patches: int) -> Tensor:
+        return self.forecast_query.view(1, 1, -1).expand(rows, n_patches, -1)
+
+
+def monotone_quantiles(raw: Tensor, levels: Tuple[float, ...], width_factor: float) -> Tensor:
+    m = levels.index(0.5)
+    widths = [0.0] * len(levels)
+    for i in range(m + 1, len(levels)):
+        widths[i] = levels[i] - levels[i - 1]
+    for i in range(m):
+        widths[i] = levels[i + 1] - levels[i]
+    widths_t = torch.tensor(widths, device=raw.device, dtype=raw.dtype)
+    inc = width_factor * widths_t * F.softplus(raw)
+    median = raw[..., m : m + 1]
+    up = torch.cumsum(inc[..., m + 1 :], dim=-1)
+    down = torch.flip(torch.cumsum(torch.flip(inc[..., :m], dims=[-1]), dim=-1), dims=[-1])
+    big = torch.finfo(raw.dtype).max
+    return torch.cat([median - down, median, median + up], dim=-1).clamp(min=-big, max=big)
+
+
+class QuantileHead(nn.Module):
+    def __init__(self, config: TelosConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.patch_size = config.patch_size
+        self.num_quantiles = config.num_quantiles
+        self.projection = ResidualBlock(
+            config.d_model,
+            config.head_hidden_dim,
+            config.patch_size * config.num_quantiles,
+            use_bias=False,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        raw = self.projection(x)
+        *leading, n_patches, _ = raw.shape
+        preds = raw.reshape(*leading, n_patches, self.patch_size, self.num_quantiles)
+        preds = preds.to(torch.promote_types(preds.dtype, torch.float32))
+        preds = monotone_quantiles(preds, self.config.quantile_knots, self.config.quantile_width_factor)
+        return preds.reshape(*leading, n_patches * self.patch_size, self.num_quantiles)
+
+
+# ==============================================================================
+# 4. Complete Standalone Model & Block-AR Forecaster
+# ==============================================================================
+
+
+class JEPALatentPredictor(nn.Module):
+    """Horizon-conditioned JEPA latent predictor $P_\\phi(\\mathbf{z}^{\\text{ctx}}_{1:L}, \\mathbf{q}_{L+1:L+H}) \\to \\hat{\\mathbf{z}}_{L+1:L+H}$.
+
+    Takes causal context representations from the encoder stream plus horizon query states and predicts
+    the future target encoder representations in latent space using RoPE-conditioned gated attention + SwiGLU blocks.
+    """
+
+    def __init__(self, config: TelosConfig) -> None:
+        super().__init__()
+        c = config
+        self.max_horizon_patches = 64
+        self.horizon_step_emb = nn.Parameter(torch.randn(self.max_horizon_patches, c.d_model) * 0.02)
+        self.layers = nn.ModuleList(
+            [
+                TransformerBlock(
+                    kind="time",
+                    embed_dim=c.d_model,
+                    num_heads=c.num_heads,
+                    mlp_hidden_dim=c.jepa_predictor_hidden,
+                    norm_eps=c.norm_eps,
+                    attn_bias=c.attn_bias,
+                    mlp_bias=c.mlp_bias,
+                    time_attn_gate=c.time_attn_gate,
+                    variate_attn_gate=c.variate_attn_gate,
+                )
+                for _ in range(c.jepa_predictor_layers)
+            ]
+        )
+        self.rope = SinCosRoPE(dim=c.head_dim, base=c.rope_theta, rotary_dim=c.rope_dim)
+        self.out_norm = RMSNorm(c.d_model, c.norm_eps)
+
+    def forward(
+        self, h_ctx: Tensor, h_dec: Tensor, ctx_key_valid: Tensor, first_obs: Tensor
+    ) -> Tensor:
+        ll = h_ctx.shape[1]
+        h = h_dec.shape[1]
+        step_idx = torch.arange(h, device=h_dec.device).clamp_max(self.max_horizon_patches - 1)
+        q_dec = h_dec + self.horizon_step_emb[step_idx].to(h_dec.dtype)[None, :, :]
+        x = torch.cat([h_ctx, q_dec], dim=1)
+        time_mask = build_forecast_time_mask(ll, h, ctx_key_valid, h_ctx.device)
+        mask = TwoStreamMask(mask=time_mask, enc_len=ll)
+        positions = build_forecast_positions(ll, h, first_obs)
+        cos, sin = self.rope(x, positions)
+        for layer in self.layers:
+            x, _ = layer(x, mask, None, cos, sin)
+        return self.out_norm(x[:, ll:, :])
+
+    def forward_tiled(
+        self,
+        h_ctx: Tensor,
+        h_dec: Tensor,
+        tiled_mask: TwoStreamMask,
+        positions: Tensor,
+        step_offsets: Tensor,
+    ) -> Tensor:
+        """Tiled Multi-Origin DHF forward pass: predicts future latents at ALL tiled blocks simultaneously."""
+        step_clamped = step_offsets.clamp(0, self.max_horizon_patches - 1)
+        q_dec = h_dec + self.horizon_step_emb[step_clamped].to(h_dec.dtype)
+        x = torch.cat([h_ctx, q_dec], dim=1)
+        cos, sin = self.rope(x, positions)
+        for layer in self.layers:
+            x, _ = layer(x, tiled_mask, None, cos, sin)
+        return self.out_norm(x[:, h_ctx.shape[1] :, :])
+
+
+class TelosModel(PreTrainedModel):
+    """Telos-Small (50.1M) Standalone Time-Series JEPA Foundation Model."""
+
+    config_class = TelosConfig
+    base_model_prefix = "telos"
+
+    def __init__(self, config: TelosConfig) -> None:
+        super().__init__(config)
+        self.config = config
+        c = config
+        self.representation = PatchRepresentation(c.patch_size, c.d_model, c.patch_proj_hidden)
+        self.transformer = Transformer(c)
+        self.jepa_predictor: Optional[JEPALatentPredictor] = (
+            JEPALatentPredictor(c) if c.jepa_predictor_layers > 0 else None
+        )
+        self.head = QuantileHead(c)
+        self.post_init()
+
+    def _init_weights(self, module: nn.Module) -> None:
+        if isinstance(module, SinCosRoPE) and module.inv_freq.is_meta:
+            module.inv_freq = module._build_inv_freq(None)
+
+    def count_parameters(self) -> int:
+        return sum(p.numel() for p in self.parameters())
+
+    def scale(self, inputs: Tensor, observed: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
+        c = self.config
+        loc, scale = causal_patch_loc_scale(
+            inputs.unsqueeze(1), observed.unsqueeze(1), c.patch_size, c.minimum_scale, c.stabilize_with_global
+        )
+        loc, scale = loc.squeeze(1), scale.squeeze(1)
+        scaled = apply_affine_asinh_masked(inputs, loc, scale, observed, use_asinh=c.use_asinh)
+        return scaled, loc, scale
+
+    def _known_cov_scaling(
+        self, inputs: Tensor, observed: Tensor, scaled: Tensor, loc: Tensor, scale: Tensor, known_rows: Tensor
+    ) -> Tuple[Tensor, Tensor, Tensor]:
+        """Normalize known-future covariate rows with non-causal whole-row statistics."""
+        if not bool(known_rows.any()):
+            return scaled, loc, scale
+        loc_c, scale_c, is_binary = instance_norm_stats(inputs, observed, minimum_scale=self.config.minimum_scale)
+        known_scaled = apply_affine_asinh_masked(inputs, loc_c, scale_c, observed, use_asinh=False)
+        if self.config.use_asinh:
+            work = torch.promote_types(known_scaled.dtype, torch.float32)
+            squashed = known_scaled.to(work).asinh().to(known_scaled.dtype)
+            known_scaled = torch.where(observed & ~is_binary, squashed, known_scaled)
+        sel = known_rows[:, None]
+        return torch.where(sel, known_scaled, scaled), torch.where(sel, loc_c, loc), torch.where(sel, scale_c, scale)
+
+    def encode_context_only(
+        self,
+        inputs: Tensor,
+        observed: Tensor,
+        group_ids: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+    ) -> Tensor | Tuple[Tensor, List[Tensor]]:
+        """Pure causal Context / Target Encoder $E_\\theta(\\mathbf{x}_{1:N})$ with zero horizon query tokens.
+
+        Used by the Momentum EMA Target Encoder during V-JEPA training to compute causal target patch
+        representations $\\mathbf{z}^{\\text{tgt}}_{L+1:L+H} = E_{\\bar\\theta}([\\mathbf{x}_{1:L}, \\mathbf{y}_{L+1:L+H}])[:, L:]$.
+        """
+        c = self.config
+        ps = c.patch_size
+        rows, t = inputs.shape
+        if group_ids is None:
+            group_ids = torch.arange(rows, device=inputs.device)
+        n = t // ps
+        scaled, _, _ = self.scale(inputs, observed)
+        x_enc = self.representation(scaled, observed)
+        obs_per_patch = observed.unflatten(-1, (n, ps)).any(dim=-1)
+        key_valid = compute_enc_key_validity(obs_per_patch, c.attend_missing_patches)
+        time_mask = enc_causal_mask(key_valid).unsqueeze(1)
+        variate_mask = build_two_stream_variate_mask(group_ids, key_valid, 0)
+        unique_ids, inverse = torch.unique(group_ids, return_inverse=True)
+        first_obs = group_first_observed_patch(obs_per_patch, inverse, unique_ids.numel())[inverse]
+        positions = torch.arange(n, device=inputs.device)[None, :] - first_obs[:, None]
+        return self.transformer(x_enc, time_mask, variate_mask, positions, return_layer_states=return_layer_states)
+
+    def encode_tiled_dhf(
+        self,
+        full_inputs: Tensor,
+        full_observed: Tensor,
+        min_context: Tensor,
+        block_size: int,
+        group_ids: Optional[Tensor] = None,
+        role_ids: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, List[Tensor]]:
+        """Multi-Origin Tiled Decoupled-Horizon JEPA (`Tiled-DHF JEPA`):
+
+        Encodes causal context tokens `v_0 .. v_{N-1}` once while simultaneously predicting future latents
+        across consecutive tiled blocks of length `block_size` starting at `min_context` (`s_rows`), so a
+        single forward pass supervises `N - min_context` forecast patches across `(N - s) / block_size` origins!
+
+        Returns `(h_pred_dec (R, N, D), anchor (R, N), active (R, N), loc_patch (R, N), dec_layer_states)`.
+        """
+        c = self.config
+        ps = c.patch_size
+        rows, w = full_inputs.shape
+        n = w // ps
+        if group_ids is None:
+            group_ids = torch.arange(rows, device=full_inputs.device)
+
+        has_cov = role_ids is not None and c.covariate_enabled and not bool((role_ids == ROLE_TARGET).all())
+        target_rows = (role_ids == ROLE_TARGET) if has_cov else None
+        known_rows = (role_ids == ROLE_KNOWN_COV) if has_cov else None
+
+        scaled, loc, scale = self.scale(full_inputs, full_observed)
+        if known_rows is not None:
+            scaled, loc, scale = self._known_cov_scaling(full_inputs, full_observed, scaled, loc, scale, known_rows)
+
+        x_enc = self.representation(scaled, full_observed)
+        x_dec = self.representation.forecast_tokens(rows, n)
+
+        obs_per_patch = full_observed.unflatten(-1, (n, ps)).any(dim=-1)
+        key_valid = compute_enc_key_validity(obs_per_patch, c.attend_missing_patches)
+        anchor, active = compute_block_layout(min_context, n, block_size)
+        time_mask = build_two_stream_time_mask(min_context, n, block_size, key_valid, bidirectional_rows=known_rows)
+        variate_mask = build_two_stream_variate_mask(
+            group_ids, key_valid, n, target_rows=target_rows, bidirectional_rows=known_rows
+        )
+        unique_ids, inverse = torch.unique(group_ids, return_inverse=True)
+        first_obs = group_first_observed_patch(obs_per_patch, inverse, unique_ids.numel())[inverse]
+        positions = build_two_stream_positions(n, first_obs)
+        ts_mask = TwoStreamMask(mask=time_mask, enc_len=n)
+
+        x = torch.cat([x_enc, x_dec], dim=1)
+        if return_layer_states:
+            h_all, l_states = self.transformer(x, ts_mask, variate_mask, positions, return_layer_states=True)
+            dec_layers = [ls[:, n:, :] for ls in l_states]
+        else:
+            h_all = self.transformer(x, ts_mask, variate_mask, positions)
+            dec_layers = []
+
+        h_ctx, h_dec = h_all[:, :n, :], h_all[:, n:, :]
+        if self.jepa_predictor is not None:
+            p_idx = torch.arange(n, device=full_inputs.device)[None, :]
+            step_offsets = (p_idx - anchor).clamp(min=0)
+            pred_time_mask = build_two_stream_time_mask(min_context, n, block_size, key_valid, bidirectional_rows=None)
+            h_dec = self.jepa_predictor.forward_tiled(
+                h_ctx, h_dec, TwoStreamMask(mask=pred_time_mask, enc_len=n), positions, step_offsets
+            )
+
+        loc_p = loc.unflatten(-1, (n, ps))[..., 0]
+        scale_p = scale.unflatten(-1, (n, ps))[..., 0]
+        return h_dec, anchor, active, torch.stack([loc_p, scale_p], dim=-1), dec_layers
+
+    def encode_penultimate(
+        self,
+        inputs: Tensor,
+        observed: Tensor,
+        horizon_patches: int,
+        group_ids: Optional[Tensor] = None,
+        role_ids: Optional[Tensor] = None,
+        known_future: Optional[Tensor] = None,
+        known_future_mask: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+    ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor] | Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, List[Tensor]]:
+        """Run Causal Patch Scaling + Two-Stream Transformer + Horizon-Conditioned JEPA Latent Predictor.
+
+        Supports both context-only (`role_ids=None`) and covariate cross-attention (`ROLE_TARGET`,
+        `ROLE_PAST_COV`, `ROLE_KNOWN_COV` with `known_future` values in horizon patches).
+        """
+        c = self.config
+        ps = c.patch_size
+        rows, t = inputs.shape
+        if group_ids is None:
+            group_ids = torch.arange(rows, device=inputs.device)
+        ll, h = t // ps, horizon_patches
+
+        has_cov = (
+            role_ids is not None
+            and c.covariate_enabled
+            and not bool((role_ids == ROLE_TARGET).all())
+        )
+
+        if not has_cov:
+            scaled, loc, scale = self.scale(inputs, observed)
+            x_enc = self.representation(scaled, observed)
+            x_dec = self.representation.forecast_tokens(rows, h)
+
+            obs_per_patch = observed.unflatten(-1, (ll, ps)).any(dim=-1)
+            key_valid = compute_enc_key_validity(obs_per_patch, c.attend_missing_patches)
+            time_mask = build_forecast_time_mask(ll, h, key_valid, inputs.device)
+            variate_mask = build_two_stream_variate_mask(group_ids, key_valid, h)
+            unique_ids, inverse = torch.unique(group_ids, return_inverse=True)
+            first_obs = group_first_observed_patch(obs_per_patch, inverse, unique_ids.numel())[inverse]
+            positions = build_forecast_positions(ll, h, first_obs)
+
+            x = torch.cat([x_enc, x_dec], dim=1)
+            mask: Tensor | TwoStreamMask = TwoStreamMask(mask=time_mask, enc_len=ll) if c.split_context_horizon_attn else time_mask
+            if return_layer_states:
+                h_all, l_states = self.transformer(x, mask, variate_mask, positions, return_layer_states=True)
+                dec_layer_states = [ls[:, ll:, :] for ls in l_states]
+            else:
+                h_all = self.transformer(x, mask, variate_mask, positions)
+                dec_layer_states = []
+            h_raw_dec = h_all[:, ll:, :]  # (R, H_patches, d_model)
+            if self.jepa_predictor is not None:
+                h_raw_dec = self.jepa_predictor(h_all[:, :ll, :], h_raw_dec, key_valid, first_obs)
+            h_aligned, z_rot = h_raw_dec, h_raw_dec
+
+            anchor_loc = loc.unflatten(-1, (ll, ps))[..., 0][:, ll - 1 : ll]
+            anchor_scale = scale.unflatten(-1, (ll, ps))[..., 0][:, ll - 1 : ll]
+            if return_layer_states:
+                return h_aligned, z_rot, h_raw_dec, anchor_loc, anchor_scale, dec_layer_states
+            return h_aligned, z_rot, h_raw_dec, anchor_loc, anchor_scale
+
+        # Two-stream covariate window path (Enc over context + horizon, Dec over context + horizon)
+        assert role_ids is not None
+        horizon_len = h * ps
+        if known_future is None:
+            known_future = torch.zeros(rows, horizon_len, device=inputs.device, dtype=inputs.dtype)
+            known_future_mask = torch.zeros(rows, horizon_len, device=inputs.device, dtype=torch.bool)
+        assert known_future_mask is not None
+        full_inputs = torch.cat([inputs, known_future], dim=-1)
+        full_observed = torch.cat([observed, known_future_mask], dim=-1)
+        n = ll + h
+        target_rows = role_ids == ROLE_TARGET
+        known_rows = role_ids == ROLE_KNOWN_COV
+
+        scaled, loc, scale = self.scale(full_inputs, full_observed)
+        scaled, loc, scale = self._known_cov_scaling(full_inputs, full_observed, scaled, loc, scale, known_rows)
+        x_enc = self.representation(scaled, full_observed)
+        x_dec = self.representation.forecast_tokens(rows, n)
+
+        obs_per_patch = full_observed.unflatten(-1, (n, ps)).any(dim=-1)
+        key_valid = compute_enc_key_validity(obs_per_patch, c.attend_missing_patches)
+        s_rows = torch.full((rows,), ll, device=full_inputs.device, dtype=torch.long)
+        time_mask = build_two_stream_time_mask(s_rows, n, c.block_size, key_valid, bidirectional_rows=known_rows)
+        variate_mask = build_two_stream_variate_mask(
+            group_ids, key_valid, n, target_rows=target_rows, bidirectional_rows=known_rows
+        )
+        unique_ids, inverse = torch.unique(group_ids, return_inverse=True)
+        first_obs = group_first_observed_patch(obs_per_patch, inverse, unique_ids.numel())[inverse]
+        positions = build_two_stream_positions(n, first_obs)
+
+        x = torch.cat([x_enc, x_dec], dim=1)
+        mask = TwoStreamMask(mask=time_mask, enc_len=n) if c.split_context_horizon_attn else time_mask
+        if return_layer_states:
+            h_all, l_states = self.transformer(x, mask, variate_mask, positions, return_layer_states=True)
+            dec_layer_states = [ls[:, n + ll : 2 * n, :] for ls in l_states]
+        else:
+            h_all = self.transformer(x, mask, variate_mask, positions)
+            dec_layer_states = []
+        h_raw_dec = h_all[:, n + ll : 2 * n, :]  # (R, H_patches, d_model)
+        if self.jepa_predictor is not None:
+            h_raw_dec = self.jepa_predictor(h_all[:, :ll, :], h_raw_dec, key_valid[:, :ll], first_obs)
+        h_aligned, z_rot = h_raw_dec, h_raw_dec
+
+        anchor_loc = loc.unflatten(-1, (n, ps))[..., 0][:, ll - 1 : ll]
+        anchor_scale = scale.unflatten(-1, (n, ps))[..., 0][:, ll - 1 : ll]
+        if return_layer_states:
+            return h_aligned, z_rot, h_raw_dec, anchor_loc, anchor_scale, dec_layer_states
+        return h_aligned, z_rot, h_raw_dec, anchor_loc, anchor_scale
+
+    def predict_block(
+        self,
+        inputs: Tensor,
+        observed: Tensor,
+        group_ids: Tensor,
+        horizon_patches: int,
+        role_ids: Optional[Tensor] = None,
+        known_future: Optional[Tensor] = None,
+        known_future_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        h_aligned, _, _, anchor_loc, anchor_scale = self.encode_penultimate(
+            inputs, observed, horizon_patches, group_ids, role_ids, known_future, known_future_mask
+        )
+        preds_norm = self.head(h_aligned)
+        delta = getattr(self, "readout_delta", None)
+        if delta is not None and delta.shape[0] == preds_norm.shape[0]:
+            corr = torch.einsum("rhd,rdp->rhp", h_aligned, delta.to(h_aligned.dtype)).reshape(preds_norm.shape[0], -1)
+            preds_norm = preds_norm + corr[..., None].to(preds_norm.dtype)
+        return denormalize_quantiles(preds_norm, anchor_loc.unsqueeze(-1), anchor_scale.unsqueeze(-1), self.config.use_asinh)
+
+    def _refine_block(
+        self, quantiles: Tensor, context: Tensor, mask: Tensor, target_rows: Optional[Tensor] = None
+    ) -> Tensor:
+        ps = self.config.patch_size
+        n_pred = quantiles.shape[1] // ps
+        if n_pred <= 1:
+            return quantiles
+        _, loc, scale = self.scale(context, mask)
+        loc0, scale0 = loc[:, -1], scale[:, -1]
+        med_idx = self.config.median_index
+        median = quantiles[..., med_idx].unflatten(-1, (n_pred, ps)).to(torch.float64)
+        sinh_median = (median - loc0[:, None, None]) / scale0[:, None, None]
+        loc_j, scale_j = advancing_block_frames(context, mask, sinh_median, self.config.minimum_scale)
+        first = torch.zeros(loc_j.shape, dtype=torch.bool, device=loc_j.device)
+        first[:, 0] = True
+        loc_j = torch.where(first, loc0[:, None], loc_j.to(quantiles.dtype)).repeat_interleave(ps, dim=-1)
+        scale_j = torch.where(first, scale0[:, None], scale_j.to(quantiles.dtype)).repeat_interleave(ps, dim=-1)
+        sinh_norm = (quantiles - loc0[:, None, None]) / scale0[:, None, None]
+        refined = clamp_nonfinite(sinh_norm * scale_j[..., None] + loc_j[..., None])
+        if target_rows is not None:
+            refined = torch.where(target_rows[:, None, None], refined, quantiles)
+        return refined
+
+    def _assemble_rows(
+        self,
+        inputs: Tensor,
+        observation_mask: Tensor,
+        prediction_length: int,
+        past_covariates: Optional[Tensor],
+        known_covariates: Optional[Tensor],
+    ) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor], Optional[Tensor], int]:
+        ps = self.config.patch_size
+        device = inputs.device
+        b, n_tgt, ctx0 = inputs.shape
+        horizon = math.ceil(prediction_length / ps) * ps
+
+        ctx_rows = [ffill_with_mask(inputs, observation_mask)]
+        ctx_masks = [observation_mask]
+        roles = [ROLE_TARGET] * n_tgt
+
+        n_past = 0
+        if past_covariates is not None and past_covariates.shape[1] > 0:
+            n_past = past_covariates.shape[1]
+            ctx_masks.append(torch.isfinite(past_covariates))
+            ctx_rows.append(torch.nan_to_num(past_covariates, nan=0.0))
+            roles += [ROLE_PAST_COV] * n_past
+
+        n_known = 0
+        known_future_vals: Optional[Tensor] = None
+        known_future_obs: Optional[Tensor] = None
+        if known_covariates is not None and known_covariates.shape[1] > 0:
+            n_known = known_covariates.shape[1]
+            kc_ctx, kc_fut = known_covariates[..., :-prediction_length], known_covariates[..., -prediction_length:]
+            ctx_masks.append(torch.isfinite(kc_ctx))
+            ctx_rows.append(torch.nan_to_num(kc_ctx, nan=0.0))
+            roles += [ROLE_KNOWN_COV] * n_known
+            known_future_obs = torch.isfinite(kc_fut)
+            known_future_vals = torch.nan_to_num(kc_fut, nan=0.0)
+
+        rows = torch.cat(ctx_rows, dim=1)
+        mask = torch.cat(ctx_masks, dim=1)
+        if rows.shape[-1] > self.config.context_length:
+            rows = rows[:, :, -self.config.context_length :]
+            mask = mask[:, :, -self.config.context_length :]
+        rows, mask = pad_to_patch(rows, mask, ps)
+        v_all = rows.shape[1]
+        if n_past == 0 and n_known == 0:
+            return rows, mask, None, None, None, n_tgt
+
+        role_ids = torch.tensor(roles, dtype=torch.long, device=device).repeat(b)
+        known_future: Optional[Tensor] = None
+        known_future_mask: Optional[Tensor] = None
+        if n_known > 0 and known_future_vals is not None and known_future_obs is not None:
+            k0 = n_tgt + n_past
+            plen = known_future_vals.shape[-1]
+            kf = torch.zeros(b, v_all, horizon, device=device, dtype=rows.dtype)
+            kfm = torch.zeros(b, v_all, horizon, device=device, dtype=torch.bool)
+            kf[:, k0 : k0 + n_known, :plen] = known_future_vals
+            kfm[:, k0 : k0 + n_known, :plen] = known_future_obs
+            known_future = kf.reshape(b * v_all, horizon)
+            known_future_mask = kfm.reshape(b * v_all, horizon)
+        return rows, mask, role_ids, known_future, known_future_mask, n_tgt
+
+    def _forecast_raw(
+        self,
+        inputs: Tensor,
+        mask: Tensor,
+        prediction_length: int,
+        past_covariates: Optional[Tensor] = None,
+        known_covariates: Optional[Tensor] = None,
+    ) -> Tensor:
+        """`(B, V, T)` -> `(B, V, H, Q_21)` with block-AR decoding, covariates, and Welford frame refinement."""
+        self.eval()
+        ps = self.config.patch_size
+        has_variate_op = any(kind != "time" for kind in self.config.block_schedule)
+        if not has_variate_op:
+            past_covariates = None
+            known_covariates = None
+
+        rows, obs_mask, role_ids, known_future, known_future_mask, n_tgt = self._assemble_rows(
+            inputs, mask, prediction_length, past_covariates, known_covariates
+        )
+        b, v, ctx_len = rows.shape
+        context = rows.reshape(b * v, ctx_len)
+        obs = obs_mask.reshape(b * v, ctx_len)
+        group_ids = torch.arange(b, device=inputs.device).repeat_interleave(v)
+
+        known_rows = past_rows = target_rows = None
+        if role_ids is not None:
+            known_rows = role_ids == ROLE_KNOWN_COV
+            past_rows = (role_ids != ROLE_TARGET) & ~known_rows
+            target_rows = role_ids == ROLE_TARGET
+
+        # Compute context min/max and last-patch scale for `quantile_real_cap_k` clamping on target rows
+        tgt_rows_3d = rows[:, :n_tgt]
+        tgt_mask_3d = obs_mask[:, :n_tgt]
+        not_obs = ~tgt_mask_3d
+        ctx_max = torch.nan_to_num(tgt_rows_3d.masked_fill(not_obs, float("-inf")).amax(-1, keepdim=True), neginf=0.0)
+        ctx_min = torch.nan_to_num(tgt_rows_3d.masked_fill(not_obs, float("inf")).amin(-1, keepdim=True), posinf=0.0)
+        _, scale0 = causal_patch_loc_scale(
+            tgt_rows_3d, tgt_mask_3d, ps, self.config.minimum_scale, self.config.stabilize_with_global
+        )
+        anchor_scale = scale0[..., -1:] * float(self.config.quantile_real_cap_k)
+
+        blocks: List[Tensor] = []
+        remaining = math.ceil(prediction_length / ps)
+        max_block = min(self.config.block_size, max(1, int(self.config.per_shot_horizon_patches)))
+        med_idx = self.config.median_index
+        offset = 0
+
+        while remaining > 0:
+            blk = min(max_block, remaining)
+            blen = blk * ps
+            kf_blk = known_future[:, offset : offset + blen] if known_future is not None else None
+            kfm_blk = known_future_mask[:, offset : offset + blen] if known_future_mask is not None else None
+            q_blk = self.predict_block(context, obs, group_ids, blk, role_ids, kf_blk, kfm_blk)
+            q_blk = self._refine_block(q_blk, context, obs, target_rows)
+            blocks.append(q_blk)
+            remaining -= blk
+            offset += blen
+            if remaining <= 0:
+                break
+            feedback = torch.nan_to_num(q_blk[..., med_idx], nan=0.0, posinf=0.0, neginf=0.0)
+            fb_mask = torch.ones_like(feedback, dtype=torch.bool)
+            if role_ids is not None and known_rows is not None and past_rows is not None:
+                if kf_blk is not None and kfm_blk is not None:
+                    feedback = torch.where(known_rows[:, None], kf_blk, feedback)
+                    fb_mask = torch.where(known_rows[:, None], kfm_blk, fb_mask)
+                feedback = torch.where(past_rows[:, None], torch.zeros_like(feedback), feedback)
+                fb_mask = fb_mask & ~past_rows[:, None]
+            context = torch.cat([context, feedback], dim=-1)
+            obs = torch.cat([obs, fb_mask], dim=-1)
+            if context.shape[-1] > self.config.context_length:
+                keep = max(ps, (self.config.context_length // ps) * ps)
+                context = context[..., -keep:]
+                obs = obs[..., -keep:]
+
+        packed = torch.cat(blocks, dim=-2)[..., :prediction_length, :]
+        quantiles = packed.reshape(b, v, prediction_length, -1)[:, :n_tgt]
+        return quantiles.clamp((ctx_min - anchor_scale).unsqueeze(-1), (ctx_max + anchor_scale).unsqueeze(-1))
+
+    @torch.inference_mode()
+    def forecast(
+        self,
+        context: np.ndarray | Tensor,
+        prediction_length: int,
+        past_covariates: Optional[np.ndarray | Tensor] = None,
+        known_covariates: Optional[np.ndarray | Tensor] = None,
+        symmetric_averaging: bool = True,
+        clip_nonnegative: bool = True,
+    ) -> Tensor:
+        """Full zero-shot probabilistic forecast `(B, V, H, 21)` with covariates & reflection averaging."""
+        device = next(self.parameters()).device
+        x = torch.as_tensor(context, dtype=torch.float32, device=device)
+        if x.dim() == 1:
+            x = x[None, None]
+        elif x.dim() == 2:
+            x = x[:, None]
+        mask = torch.isfinite(x)
+        inputs = torch.where(mask, x, torch.zeros_like(x))
+
+        pc_t: Optional[Tensor] = None
+        if past_covariates is not None:
+            pc_t = torch.as_tensor(past_covariates, dtype=torch.float32, device=device)
+            if pc_t.dim() == 2:
+                pc_t = pc_t[:, None]
+
+        kc_t: Optional[Tensor] = None
+        if known_covariates is not None:
+            kc_t = torch.as_tensor(known_covariates, dtype=torch.float32, device=device)
+            if kc_t.dim() == 2:
+                kc_t = kc_t[:, None]
+
+        def _single_view(inp: Tensor, msk: Tensor, pc: Optional[Tensor], kc: Optional[Tensor]) -> Tensor:
+            q = self._forecast_raw(inp, msk, prediction_length, pc, kc)
+            if symmetric_averaging:
+                mir = self._forecast_raw(
+                    -inp,
+                    msk,
+                    prediction_length,
+                    -pc if pc is not None else None,
+                    -kc if kc is not None else None,
+                )
+                q = 0.5 * (q + (-mir).flip(-1))
+            return q
+
+        quantiles = _single_view(inputs, mask, pc_t, kc_t)
+        if clip_nonnegative:
+            has_obs = mask.any(dim=-1)
+            min_obs = torch.where(mask, inputs, torch.full_like(inputs, float("inf"))).amin(dim=-1)
+            eligible = has_obs & (min_obs >= 0.0)
+            quantiles = torch.where(eligible[..., None, None], quantiles.clamp_min(0.0), quantiles)
+        return quantiles
+
+    @torch.inference_mode()
+    def predict_series_batch(
+        self,
+        series_list: List[np.ndarray],
+        horizon: int,
+        quantile_levels: List[float],
+        batch_size: int = 128,
+        past_covariates_list: Optional[List[Optional[np.ndarray]]] = None,
+        known_covariates_list: Optional[List[Optional[np.ndarray]]] = None,
+        max_covariates: int = 4,
+    ) -> np.ndarray:
+        """Standalone batch inference for arbitrary-length 1D numpy series + optional covariates.
+
+        Returns `(N, horizon, 1 + len(quantile_levels))` where col 0 is median (0.5) and
+        cols `1..` correspond to `quantile_levels`.
+        """
+        self.eval()
+        N = len(series_list)
+        out = np.zeros((N, horizon, 1 + len(quantile_levels)), dtype=np.float64)
+        knots = np.asarray(self.config.quantile_knots, dtype=np.float64)
+        med_idx = self.config.median_index
+
+        has_variate_op = any(kind != "time" for kind in self.config.block_schedule)
+        use_cov = has_variate_op and (past_covariates_list is not None or known_covariates_list is not None)
+
+        # Group by similar lengths (and covariate channel counts when present) for efficient padding
+        def _sort_key(i: int) -> Tuple[int, int, int]:
+            c_p = 0
+            if use_cov and past_covariates_list is not None and past_covariates_list[i] is not None:
+                c_p = min(past_covariates_list[i].shape[1], max_covariates)
+            c_k = 0
+            if use_cov and known_covariates_list is not None and known_covariates_list[i] is not None:
+                c_k = min(known_covariates_list[i].shape[1], max_covariates)
+            return (c_p, c_k, len(series_list[i]))
+
+        order = sorted(range(N), key=_sort_key)
+        # Covariate rows share dense variate attention across the batch, so cap rows per forward pass.
+        step = max(8, batch_size // 8) if use_cov else batch_size
+        for start in range(0, N, step):
+            idx = order[start : start + step]
+            max_t = min(self.config.context_length, max(len(series_list[i]) for i in idx))
+            B_cur = len(idx)
+            batch_arr = np.full((B_cur, 1, max_t), np.nan, dtype=np.float32)
+
+            c_p_max = 0
+            c_k_max = 0
+            if use_cov:
+                if past_covariates_list is not None:
+                    c_p_max = min(max_covariates, max((past_covariates_list[i].shape[1] if past_covariates_list[i] is not None else 0) for i in idx))
+                if known_covariates_list is not None:
+                    c_k_max = min(max_covariates, max((known_covariates_list[i].shape[1] if known_covariates_list[i] is not None else 0) for i in idx))
+
+            pc_arr = np.full((B_cur, c_p_max, max_t), np.nan, dtype=np.float32) if c_p_max > 0 else None
+            kc_arr = np.full((B_cur, c_k_max, max_t + horizon), np.nan, dtype=np.float32) if c_k_max > 0 else None
+
+            for b_i, orig_i in enumerate(idx):
+                y = np.asarray(series_list[orig_i], dtype=np.float32)
+                if len(y) > max_t:
+                    y = y[-max_t:]
+                L_i = len(y)
+                batch_arr[b_i, 0, max_t - L_i :] = y
+
+                if pc_arr is not None and past_covariates_list is not None and past_covariates_list[orig_i] is not None:
+                    xp = np.asarray(past_covariates_list[orig_i], dtype=np.float32)
+                    if xp.shape[0] > L_i:
+                        xp = xp[-L_i:]
+                    nc = min(xp.shape[1], c_p_max)
+                    pc_arr[b_i, :nc, max_t - xp.shape[0] :] = xp[:, :nc].T
+
+                if kc_arr is not None and known_covariates_list is not None and known_covariates_list[orig_i] is not None:
+                    xk = np.asarray(known_covariates_list[orig_i], dtype=np.float32)
+                    # xk has shape (T_orig + horizon, C_k)
+                    xk_past = xk[:-horizon]
+                    xk_fut = xk[-horizon:]
+                    if xk_past.shape[0] > L_i:
+                        xk_past = xk_past[-L_i:]
+                    nc = min(xk.shape[1], c_k_max)
+                    for c_j in range(nc):
+                        col_p = xk_past[:, c_j]
+                        col_f = xk_fut[:, c_j]
+                        # Target-encode low-cardinality discrete/categorical covariates (>2 distinct values, <=12)
+                        uniq = np.unique(col_p[np.isfinite(col_p)])
+                        if 2 < len(uniq) <= 12 and np.all(np.floor(uniq) == uniq):
+                            col_p, col_f_enc = target_encode_covariate(y[-len(col_p):], col_p, col_f)
+                            if col_f_enc is not None:
+                                col_f = col_f_enc
+                        kc_arr[b_i, c_j, max_t - len(col_p) : max_t] = col_p
+                        kc_arr[b_i, c_j, max_t : max_t + horizon] = col_f
+
+            q21 = self.forecast(batch_arr, horizon, past_covariates=pc_arr, known_covariates=kc_arr).double().cpu().numpy()[:, 0, :, :]
+            for b_i, orig_i in enumerate(idx):
+                out[orig_i, :, 0] = q21[b_i, :, med_idx]
+                for k_l, lv in enumerate(quantile_levels):
+                    match = [k for k, m in enumerate(knots) if abs(m - lv) < 1e-9]
+                    if match:
+                        out[orig_i, :, 1 + k_l] = q21[b_i, :, match[0]]
+                    else:
+                        for h_i in range(horizon):
+                            out[orig_i, h_i, 1 + k_l] = np.interp(lv, knots, q21[b_i, h_i])
+        return out
+
+
+
+    def forward(
+        self,
+        context: Tensor,
+        prediction_length: int = 96,
+        past_covariates: Optional[Tensor] = None,
+        known_covariates: Optional[Tensor] = None,
+        symmetric_averaging: bool = True,
+        clip_nonnegative: bool = True,
+    ) -> Tensor:
+        """Forward alias for `forecast(...)` returning `(B, V, prediction_length, 21)` quantiles."""
+        return self.forecast(
+            context=context,
+            prediction_length=prediction_length,
+            past_covariates=past_covariates,
+            known_covariates=known_covariates,
+            symmetric_averaging=symmetric_averaging,
+            clip_nonnegative=clip_nonnegative,
+        )
+
+
+TelosForForecasting = TelosModel
+
+
+def save_telos_checkpoint(model: TelosModel, path: Path | str) -> None:
+    p = Path(path)
+    p.parent.mkdir(parents=True, exist_ok=True)
+    payload = {
+        "config": model.config.to_dict(),
+        "num_parameters": model.count_parameters(),
+        "state_dict": model.state_dict(),
+    }
+    torch.save(payload, p)
+
+
+def load_telos_checkpoint(path: Path | str, device: str = "cpu") -> TelosModel:
+    payload = torch.load(Path(path), map_location=device, weights_only=False)
+    cfg_dict = dict(payload["config"])
+    cfg_dict["model_name"] = "Telos-Small"
+    cfg = TelosConfig(**cfg_dict)
+    model = TelosModel(cfg).to(device)
+    model.load_state_dict(payload["state_dict"])
+    model.eval()
+    return model

pyproject.toml

@@ -0,0 +1,17 @@
+[build-system]
+requires = ["setuptools>=64", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "telos-ts"
+version = "0.1.0"
+description = "Telos-Small (50.1M): Standalone Time-Series JEPA Foundation Model"
+readme = "README.md"
+license = { file = "LICENSE" }
+requires-python = ">=3.10"
+dependencies = [
+    "torch>=2.2.0",
+    "numpy>=1.24.0",
+    "safetensors>=0.4.0",
+    "transformers>=4.40.0",
+]

requirements.txt

@@ -0,0 +1,4 @@
+torch>=2.2.0
+numpy>=1.24.0
+safetensors>=0.4.0
+transformers>=4.40.0

results/fev_bench_telos_small_50m.json

@@ -0,0 +1,772 @@
+{
+  "benchmark": "AutoGluon FEV-Bench (`fev-bench-mini`: 20 Multi-Domain & Covariate Forecasting Tasks \u2014 100% Standalone)",
+  "tasks_evaluated": 20,
+  "checkpoint_sha256": {
+    "telos_small": "ac7b7d2ff259aabdc7d94afccafae9790d2f9d3bb86097d16235ca10f7f2f7a7"
+  },
+  "nn_context_length": 4096,
+  "method": "Blended rows: quantile average of the analytical seasonal-trend solver, a multi-cycle empirical candidate, and the network (with past/known covariates as extra rows; plus a covariate-free network + analytical covariate effect candidate on covariate tasks). Convex weights per (window, target column) are fitted on historical backtests inside past_data only. NN-only rows are the raw network output with covariates and no blend.",
+  "elapsed_s": 657.3,
+  "leaderboard": [
+    {
+      "model_name": "Timestral-209M",
+      "mase_skill_pct": 46.25,
+      "sql_skill_pct": 56.19,
+      "median_inference_s": 3.177,
+      "mase_win_rate_pct": 88.08
+    },
+    {
+      "model_name": "timesfm-3",
+      "mase_skill_pct": 45.17,
+      "sql_skill_pct": 54.96,
+      "median_inference_s": 21.208,
+      "mase_win_rate_pct": 80.77
+    },
+    {
+      "model_name": "Timestral-28M",
+      "mase_skill_pct": 43.74,
+      "sql_skill_pct": 53.94,
+      "median_inference_s": 0.581,
+      "mase_win_rate_pct": 79.23
+    },
+    {
+      "model_name": "Chronos-2",
+      "mase_skill_pct": 40.36,
+      "sql_skill_pct": 51.12,
+      "median_inference_s": 1.936,
+      "mase_win_rate_pct": 66.15
+    },
+    {
+      "model_name": "TabPFN-TS-3",
+      "mase_skill_pct": 38.73,
+      "sql_skill_pct": 49.23,
+      "median_inference_s": 947.49,
+      "mase_win_rate_pct": 56.92
+    },
+    {
+      "model_name": "TimesFM-2.5",
+      "mase_skill_pct": 37.87,
+      "sql_skill_pct": 48.54,
+      "median_inference_s": 36.18,
+      "mase_win_rate_pct": 52.69
+    },
+    {
+      "model_name": "Toto-2.0-2.5B",
+      "mase_skill_pct": 33.83,
+      "sql_skill_pct": 46.07,
+      "median_inference_s": 15.762,
+      "mase_win_rate_pct": 59.23
+    },
+    {
+      "model_name": "Telos-Small (50.1M Wts)",
+      "mase_skill_pct": 31.62,
+      "sql_skill_pct": 42.82,
+      "median_inference_s": 5.929,
+      "mase_win_rate_pct": 40.0
+    },
+    {
+      "model_name": "TFT",
+      "mase_skill_pct": 25.37,
+      "sql_skill_pct": 35.26,
+      "median_inference_s": 3.555,
+      "mase_win_rate_pct": 36.35
+    },
+    {
+      "model_name": "PatchTST",
+      "mase_skill_pct": 23.85,
+      "sql_skill_pct": 34.96,
+      "median_inference_s": 6.27,
+      "mase_win_rate_pct": 28.08
+    },
+    {
+      "model_name": "CatBoost",
+      "mase_skill_pct": 30.46,
+      "sql_skill_pct": 28.41,
+      "median_inference_s": 1.84,
+      "mase_win_rate_pct": 44.23
+    },
+    {
+      "model_name": "LightGBM",
+      "mase_skill_pct": 26.16,
+      "sql_skill_pct": 23.99,
+      "median_inference_s": 1.522,
+      "mase_win_rate_pct": 37.69
+    },
+    {
+      "model_name": "Seasonal Naive",
+      "mase_skill_pct": 0.0,
+      "sql_skill_pct": 0.0,
+      "median_inference_s": 2.371,
+      "mase_win_rate_pct": 5.19
+    }
+  ],
+  "pure_nn_leaderboard": [
+    {
+      "model_name": "Telos-Small (Pure JEPA)",
+      "mase_skill_pct": 26.3,
+      "sql_skill_pct": 39.49,
+      "median_inference_s": 1.416,
+      "mase_win_rate_pct": 0.0
+    }
+  ],
+  "calibration_diagnostics": [
+    {
+      "dataset_config": "boomlet_1282",
+      "horizon": 60,
+      "mean_blend_weights": {
+        "large": [
+          0.118,
+          0.134,
+          0.748
+        ]
+      },
+      "backtest_series": 700,
+      "backtest_horizon": 60,
+      "covariates": false
+    },
+    {
+      "dataset_config": "boomlet_1676",
+      "horizon": 96,
+      "mean_blend_weights": {
+        "large": [
+          0.204,
+          0.166,
+          0.63
+        ]
+      },
+      "backtest_series": 2000,
+      "backtest_horizon": 96,
+      "covariates": false
+    },
+    {
+      "dataset_config": "boomlet_619",
+      "horizon": 60,
+      "mean_blend_weights": {
+        "large": [
+          0.056,
+          0.079,
+          0.864
+        ]
+      },
+      "backtest_series": 1040,
+      "backtest_horizon": 60,
+      "covariates": false
+    },
+    {
+      "dataset_config": "bizitobs_l2c_5T",
+      "horizon": 288,
+      "mean_blend_weights": {
+        "large": [
+          0.017,
+          0.785,
+          0.198
+        ]
+      },
+      "backtest_series": 140,
+      "backtest_horizon": 288,
+      "covariates": false
+    },
+    {
+      "dataset_config": "epf_np",
+      "horizon": 24,
+      "mean_blend_weights": {
+        "large": [
+          0.177,
+          0.089,
+          0.053,
+          0.682
+        ]
+      },
+      "backtest_series": 20,
+      "backtest_horizon": 24,
+      "covariates": true
+    },
+    {
+      "dataset_config": "ETT_15T",
+      "horizon": 96,
+      "mean_blend_weights": {
+        "large": [
+          0.287,
+          0.149,
+          0.565
+        ]
+      },
+      "backtest_series": 280,
+      "backtest_horizon": 96,
+      "covariates": false
+    },
+    {
+      "dataset_config": "ETT_1H",
+      "horizon": 168,
+      "mean_blend_weights": {
+        "large": [
+          0.279,
+          0.255,
+          0.466
+        ]
+      },
+      "backtest_series": 280,
+      "backtest_horizon": 168,
+      "covariates": false
+    },
+    {
+      "dataset_config": "proenfo_gfc14",
+      "horizon": 168,
+      "mean_blend_weights": {
+        "large": [
+          0.102,
+          0.177,
+          0.487,
+          0.234
+        ]
+      },
+      "backtest_series": 20,
+      "backtest_horizon": 168,
+      "covariates": true
+    },
+    {
+      "dataset_config": "hospital_admissions_1W",
+      "horizon": 13,
+      "mean_blend_weights": {
+        "large": [
+          0.021,
+          0.073,
+          0.906
+        ]
+      },
+      "backtest_series": 128,
+      "backtest_horizon": 13,
+      "covariates": false
+    },
+    {
+      "dataset_config": "hospital_admissions_1D",
+      "horizon": 28,
+      "mean_blend_weights": {
+        "large": [
+          0.05,
+          0.089,
+          0.862
+        ]
+      },
+      "backtest_series": 160,
+      "backtest_horizon": 28,
+      "covariates": false
+    },
+    {
+      "dataset_config": "jena_weather_1H",
+      "horizon": 24,
+      "mean_blend_weights": {
+        "large": [
+          0.16,
+          0.138,
+          0.702
+        ]
+      },
+      "backtest_series": 420,
+      "backtest_horizon": 24,
+      "covariates": false
+    },
+    {
+      "dataset_config": "M_DENSE_1D",
+      "horizon": 28,
+      "mean_blend_weights": {
+        "large": [
+          0.064,
+          0.199,
+          0.737
+        ]
+      },
+      "backtest_series": 160,
+      "backtest_horizon": 28,
+      "covariates": false
+    },
+    {
+      "dataset_config": "rohlik_orders_1D",
+      "horizon": 61,
+      "mean_blend_weights": {
+        "large": [
+          0.208,
+          0.492,
+          0.194,
+          0.105
+        ]
+      },
+      "backtest_series": 35,
+      "backtest_horizon": 61,
+      "covariates": true
+    },
+    {
+      "dataset_config": "rossmann_1W",
+      "horizon": 13,
+      "mean_blend_weights": {
+        "large": [
+          0.028,
+          0.217,
+          0.057,
+          0.698
+        ]
+      },
+      "backtest_series": 128,
+      "backtest_horizon": 13,
+      "covariates": true
+    },
+    {
+      "dataset_config": "rossmann_1D",
+      "horizon": 48,
+      "mean_blend_weights": {
+        "large": [
+          0.041,
+          0.556,
+          0.017,
+          0.386
+        ]
+      },
+      "backtest_series": 160,
+      "backtest_horizon": 48,
+      "covariates": true
+    },
+    {
+      "dataset_config": "solar_with_weather_1H",
+      "horizon": 24,
+      "mean_blend_weights": {
+        "large": [
+          0.317,
+          0.204,
+          0.052,
+          0.426
+        ]
+      },
+      "backtest_series": 20,
+      "backtest_horizon": 24,
+      "covariates": true
+    },
+    {
+      "dataset_config": "uci_air_quality_1H",
+      "horizon": 168,
+      "mean_blend_weights": {
+        "large": [
+          0.133,
+          0.051,
+          0.119,
+          0.697
+        ]
+      },
+      "backtest_series": 80,
+      "backtest_horizon": 168,
+      "covariates": true
+    },
+    {
+      "dataset_config": "uk_covid_nation_1D",
+      "horizon": 28,
+      "mean_blend_weights": {
+        "large": [
+          0.037,
+          0.925,
+          0.007,
+          0.031
+        ]
+      },
+      "backtest_series": 240,
+      "backtest_horizon": 28,
+      "covariates": true
+    },
+    {
+      "dataset_config": "us_consumption_1Y",
+      "horizon": 5,
+      "mean_blend_weights": {
+        "large": [
+          0.224,
+          0.123,
+          0.652
+        ]
+      },
+      "backtest_series": 128,
+      "backtest_horizon": 5,
+      "covariates": false
+    },
+    {
+      "dataset_config": "world_co2_emissions",
+      "horizon": 5,
+      "mean_blend_weights": {
+        "large": [
+          0.263,
+          0.094,
+          0.644
+        ]
+      },
+      "backtest_series": 112,
+      "backtest_horizon": 5,
+      "covariates": false
+    }
+  ],
+  "model_family": "Telos",
+  "model_name": "Telos-Small",
+  "parameters": 50100747,
+  "telos_small_parameters": 50100747,
+  "telos_small_per_task": [
+    {
+      "dataset_config": "boomlet_1282",
+      "horizon": 60,
+      "SQL": 0.4257,
+      "MASE": 0.5239
+    },
+    {
+      "dataset_config": "boomlet_1676",
+      "horizon": 96,
+      "SQL": 0.6215,
+      "MASE": 0.7731
+    },
+    {
+      "dataset_config": "boomlet_619",
+      "horizon": 60,
+      "SQL": 0.4566,
+      "MASE": 0.5799
+    },
+    {
+      "dataset_config": "bizitobs_l2c_5T",
+      "horizon": 288,
+      "SQL": 0.5054,
+      "MASE": 0.5516
+    },
+    {
+      "dataset_config": "epf_np",
+      "horizon": 24,
+      "SQL": 1.0904,
+      "MASE": 1.3333
+    },
+    {
+      "dataset_config": "ETT_15T",
+      "horizon": 96,
+      "SQL": 0.5947,
+      "MASE": 0.7548
+    },
+    {
+      "dataset_config": "ETT_1H",
+      "horizon": 168,
+      "SQL": 0.8858,
+      "MASE": 1.1242
+    },
+    {
+      "dataset_config": "proenfo_gfc14",
+      "horizon": 168,
+      "SQL": 0.733,
+      "MASE": 0.9065
+    },
+    {
+      "dataset_config": "hospital_admissions_1W",
+      "horizon": 13,
+      "SQL": 0.584,
+      "MASE": 0.7594
+    },
+    {
+      "dataset_config": "hospital_admissions_1D",
+      "horizon": 28,
+      "SQL": 0.5554,
+      "MASE": 0.7183
+    },
+    {
+      "dataset_config": "jena_weather_1H",
+      "horizon": 24,
+      "SQL": 0.3997,
+      "MASE": 0.4864
+    },
+    {
+      "dataset_config": "M_DENSE_1D",
+      "horizon": 28,
+      "SQL": 0.9012,
+      "MASE": 1.0869
+    },
+    {
+      "dataset_config": "rohlik_orders_1D",
+      "horizon": 61,
+      "SQL": 1.1232,
+      "MASE": 1.3794
+    },
+    {
+      "dataset_config": "rossmann_1W",
+      "horizon": 13,
+      "SQL": 0.4484,
+      "MASE": 0.5559
+    },
+    {
+      "dataset_config": "rossmann_1D",
+      "horizon": 48,
+      "SQL": 0.3514,
+      "MASE": 0.4346
+    },
+    {
+      "dataset_config": "solar_with_weather_1H",
+      "horizon": 24,
+      "SQL": 0.9153,
+      "MASE": 1.083
+    },
+    {
+      "dataset_config": "uci_air_quality_1H",
+      "horizon": 168,
+      "SQL": 0.8607,
+      "MASE": 1.0959
+    },
+    {
+      "dataset_config": "uk_covid_nation_1D",
+      "horizon": 28,
+      "SQL": 6.2561,
+      "MASE": 7.6166
+    },
+    {
+      "dataset_config": "us_consumption_1Y",
+      "horizon": 5,
+      "SQL": 3.9744,
+      "MASE": 4.8055
+    },
+    {
+      "dataset_config": "world_co2_emissions",
+      "horizon": 5,
+      "SQL": 2.788,
+      "MASE": 3.3475
+    }
+  ],
+  "telos_small_pure_jepa_per_task": [
+    {
+      "dataset_config": "boomlet_1282",
+      "horizon": 60,
+      "SQL": 0.4261,
+      "MASE": 0.5198
+    },
+    {
+      "dataset_config": "boomlet_1676",
+      "horizon": 96,
+      "SQL": 0.6309,
+      "MASE": 0.7968
+    },
+    {
+      "dataset_config": "boomlet_619",
+      "horizon": 60,
+      "SQL": 0.4533,
+      "MASE": 0.5775
+    },
+    {
+      "dataset_config": "bizitobs_l2c_5T",
+      "horizon": 288,
+      "SQL": 0.5816,
+      "MASE": 0.7152
+    },
+    {
+      "dataset_config": "epf_np",
+      "horizon": 24,
+      "SQL": 1.0511,
+      "MASE": 1.2815
+    },
+    {
+      "dataset_config": "ETT_15T",
+      "horizon": 96,
+      "SQL": 0.5993,
+      "MASE": 0.7625
+    },
+    {
+      "dataset_config": "ETT_1H",
+      "horizon": 168,
+      "SQL": 0.8881,
+      "MASE": 1.1321
+    },
+    {
+      "dataset_config": "proenfo_gfc14",
+      "horizon": 168,
+      "SQL": 0.7714,
+      "MASE": 0.9484
+    },
+    {
+      "dataset_config": "hospital_admissions_1W",
+      "horizon": 13,
+      "SQL": 0.5832,
+      "MASE": 0.7583
+    },
+    {
+      "dataset_config": "hospital_admissions_1D",
+      "horizon": 28,
+      "SQL": 0.5545,
+      "MASE": 0.7172
+    },
+    {
+      "dataset_config": "jena_weather_1H",
+      "horizon": 24,
+      "SQL": 0.3868,
+      "MASE": 0.473
+    },
+    {
+      "dataset_config": "M_DENSE_1D",
+      "horizon": 28,
+      "SQL": 0.8446,
+      "MASE": 1.0358
+    },
+    {
+      "dataset_config": "rohlik_orders_1D",
+      "horizon": 61,
+      "SQL": 1.1576,
+      "MASE": 1.4274
+    },
+    {
+      "dataset_config": "rossmann_1W",
+      "horizon": 13,
+      "SQL": 0.5059,
+      "MASE": 0.6728
+    },
+    {
+      "dataset_config": "rossmann_1D",
+      "horizon": 48,
+      "SQL": 0.5364,
+      "MASE": 0.6487
+    },
+    {
+      "dataset_config": "solar_with_weather_1H",
+      "horizon": 24,
+      "SQL": 0.9113,
+      "MASE": 1.1147
+    },
+    {
+      "dataset_config": "uci_air_quality_1H",
+      "horizon": 168,
+      "SQL": 0.8885,
+      "MASE": 1.1308
+    },
+    {
+      "dataset_config": "uk_covid_nation_1D",
+      "horizon": 28,
+      "SQL": 10.8811,
+      "MASE": 14.4062
+    },
+    {
+      "dataset_config": "us_consumption_1Y",
+      "horizon": 5,
+      "SQL": 3.7773,
+      "MASE": 4.696
+    },
+    {
+      "dataset_config": "world_co2_emissions",
+      "horizon": 5,
+      "SQL": 2.6709,
+      "MASE": 3.2729
+    }
+  ],
+  "analytical_baseline_per_task": [
+    {
+      "dataset_config": "boomlet_1282",
+      "horizon": 60,
+      "SQL": 0.6039,
+      "MASE": 0.7384
+    },
+    {
+      "dataset_config": "boomlet_1676",
+      "horizon": 96,
+      "SQL": 0.6771,
+      "MASE": 0.825
+    },
+    {
+      "dataset_config": "boomlet_619",
+      "horizon": 60,
+      "SQL": 0.6681,
+      "MASE": 0.8765
+    },
+    {
+      "dataset_config": "bizitobs_l2c_5T",
+      "horizon": 288,
+      "SQL": 0.6914,
+      "MASE": 0.8868
+    },
+    {
+      "dataset_config": "epf_np",
+      "horizon": 24,
+      "SQL": 1.1541,
+      "MASE": 1.386
+    },
+    {
+      "dataset_config": "ETT_15T",
+      "horizon": 96,
+      "SQL": 0.618,
+      "MASE": 0.7603
+    },
+    {
+      "dataset_config": "ETT_1H",
+      "horizon": 168,
+      "SQL": 0.9736,
+      "MASE": 1.183
+    },
+    {
+      "dataset_config": "proenfo_gfc14",
+      "horizon": 168,
+      "SQL": 0.883,
+      "MASE": 0.9937
+    },
+    {
+      "dataset_config": "hospital_admissions_1W",
+      "horizon": 13,
+      "SQL": 0.8269,
+      "MASE": 0.8035
+    },
+    {
+      "dataset_config": "hospital_admissions_1D",
+      "horizon": 28,
+      "SQL": 0.6514,
+      "MASE": 0.8265
+    },
+    {
+      "dataset_config": "jena_weather_1H",
+      "horizon": 24,
+      "SQL": 0.4968,
+      "MASE": 0.5994
+    },
+    {
+      "dataset_config": "M_DENSE_1D",
+      "horizon": 28,
+      "SQL": 0.9467,
+      "MASE": 1.1487
+    },
+    {
+      "dataset_config": "rohlik_orders_1D",
+      "horizon": 61,
+      "SQL": 1.1746,
+      "MASE": 1.4348
+    },
+    {
+      "dataset_config": "rossmann_1W",
+      "horizon": 13,
+      "SQL": 0.7566,
+      "MASE": 0.4123
+    },
+    {
+      "dataset_config": "rossmann_1D",
+      "horizon": 48,
+      "SQL": 0.5226,
+      "MASE": 0.5088
+    },
+    {
+      "dataset_config": "solar_with_weather_1H",
+      "horizon": 24,
+      "SQL": 0.9368,
+      "MASE": 0.9999
+    },
+    {
+      "dataset_config": "uci_air_quality_1H",
+      "horizon": 168,
+      "SQL": 0.9629,
+      "MASE": 1.1434
+    },
+    {
+      "dataset_config": "uk_covid_nation_1D",
+      "horizon": 28,
+      "SQL": 9.2265,
+      "MASE": 9.8765
+    },
+    {
+      "dataset_config": "us_consumption_1Y",
+      "horizon": 5,
+      "SQL": 4.2187,
+      "MASE": 4.7991
+    },
+    {
+      "dataset_config": "world_co2_emissions",
+      "horizon": 5,
+      "SQL": 2.9047,
+      "MASE": 3.3849
+    }
+  ]
+}

results/gift_eval_telos_small_50m.json

@@ -0,0 +1,266 @@
+{
+  "benchmark": "Salesforce GIFT-Eval (General Time-Series Forecasting Benchmark \u2014 97 Tasks across 7 Domains \u2014 100% Standalone)",
+  "tasks_evaluated": 97,
+  "checkpoint_sha256": {
+    "telos_small": "ac7b7d2ff259aabdc7d94afccafae9790d2f9d3bb86097d16235ca10f7f2f7a7"
+  },
+  "nn_context_length": 4096,
+  "full_panel": true,
+  "subsample": "none: every series, every variate and every official rolling window (Seasonal-Naive recomputed on the same panel)",
+  "method": "Blended rows: quantile average of the analytical seasonal-trend solver, a multi-cycle empirical candidate, and the network, with convex weights fitted per task on up to 16 historical backtests inside each `past` (t <= T). NN-only rows are the raw network output with no blend.",
+  "seasonal_naive_recomputed_vs_published_crps": {
+    "median_ratio": 1.0,
+    "within_1pct": 73,
+    "tasks": 97,
+    "per_task": {
+      "bitbrains_fast_storage/H/short": 1.0,
+      "solar/H/short": 1.0,
+      "solar/H/medium": 1.0,
+      "solar/H/long": 1.0,
+      "m_dense/D/short": 0.5168,
+      "ett2/H/short": 1.0,
+      "ett2/H/medium": 1.0,
+      "ett2/H/long": 1.0,
+      "ett2/D/short": 1.023,
+      "electricity/15T/short": 1.0,
+      "electricity/15T/medium": 1.0,
+      "electricity/15T/long": 1.0,
+      "bitbrains_rnd/H/short": 1.0018,
+      "loop_seattle/H/short": 1.0,
+      "loop_seattle/H/medium": 1.0,
+      "loop_seattle/H/long": 1.0,
+      "solar/W/short": 1.0049,
+      "bizitobs_application/10S/short": 1.0,
+      "bizitobs_application/10S/medium": 1.0,
+      "bizitobs_application/10S/long": 1.0,
+      "bizitobs_service/10S/short": 1.0,
+      "bizitobs_service/10S/medium": 1.0,
+      "bizitobs_service/10S/long": 1.0,
+      "bizitobs_l2c/5T/short": 1.0,
+      "bizitobs_l2c/5T/medium": 1.0,
+      "bizitobs_l2c/5T/long": 1.0,
+      "ett2/W/short": 3.1426,
+      "loop_seattle/5T/short": 1.0,
+      "loop_seattle/5T/medium": 1.0,
+      "loop_seattle/5T/long": 1.0,
+      "loop_seattle/D/short": 0.5584,
+      "bitbrains_rnd/5T/short": 1.003,
+      "bitbrains_rnd/5T/medium": 1.0035,
+      "bitbrains_rnd/5T/long": 1.0026,
+      "electricity/W/short": 0.7838,
+      "solar/D/short": 0.7413,
+      "ett1/H/short": 1.0,
+      "ett1/H/medium": 1.0,
+      "ett1/H/long": 1.0,
+      "temperature_rain/D/short": 0.7636,
+      "saugeen/W/short": 0.9005,
+      "covid_deaths/D/short": 1.1169,
+      "us_births/D/short": 0.2782,
+      "m4_weekly/W/short": 1.83,
+      "car_parts/M/short": 1.0,
+      "sz_taxi/H/short": 1.0,
+      "ett1/15T/short": 1.0,
+      "ett1/15T/medium": 1.0,
+      "ett1/15T/long": 1.0,
+      "saugeen/M/short": 1.0,
+      "us_births/M/short": 1.0,
+      "m4_quarterly/Q/short": 1.0,
+      "ett1/W/short": 1.3819,
+      "saugeen/D/short": 1.2366,
+      "electricity/H/short": 1.0,
+      "electricity/H/medium": 1.0,
+      "electricity/H/long": 1.0,
+      "ett1/D/short": 0.8688,
+      "bizitobs_l2c/H/short": 1.0,
+      "bizitobs_l2c/H/medium": 1.0,
+      "bizitobs_l2c/H/long": 1.0,
+      "hospital/M/short": 1.0,
+      "m4_daily/D/short": 1.2157,
+      "bitbrains_fast_storage/5T/short": 1.0,
+      "bitbrains_fast_storage/5T/medium": 1.0,
+      "bitbrains_fast_storage/5T/long": 0.9998,
+      "ett2/15T/short": 1.0,
+      "ett2/15T/medium": 1.0,
+      "ett2/15T/long": 1.0,
+      "jena_weather/10T/short": 0.9999,
+      "jena_weather/10T/medium": 0.9999,
+      "jena_weather/10T/long": 0.9999,
+      "jena_weather/D/short": 0.5704,
+      "us_births/W/short": 0.8873,
+      "m_dense/H/short": 1.0,
+      "m_dense/H/medium": 1.0,
+      "m_dense/H/long": 1.0,
+      "kdd_cup_2018/D/short": 0.7431,
+      "electricity/D/short": 0.8337,
+      "restaurant/D/short": 0.5938,
+      "jena_weather/H/short": 0.9999,
+      "jena_weather/H/medium": 0.9999,
+      "jena_weather/H/long": 0.9999,
+      "m4_monthly/M/short": 1.0,
+      "hierarchical_sales/D/short": 0.6344,
+      "m4_hourly/H/short": 1.0,
+      "m4_yearly/A/short": 1.0,
+      "hierarchical_sales/W/short": 0.7673,
+      "kdd_cup_2018/H/short": 0.9927,
+      "kdd_cup_2018/H/medium": 0.9807,
+      "kdd_cup_2018/H/long": 0.9778,
+      "sz_taxi/15T/short": 1.0,
+      "sz_taxi/15T/medium": 1.0,
+      "sz_taxi/15T/long": 1.0,
+      "solar/10T/short": 1.0,
+      "solar/10T/medium": 1.0,
+      "solar/10T/long": 1.0
+    }
+  },
+  "mean_blend_weights": {
+    "telos_small": [
+      0.145,
+      0.251,
+      0.605
+    ]
+  },
+  "elapsed_s": 5505.5,
+  "leaderboard": [
+    {
+      "model_name": "Timestral-209M",
+      "provider": "Mistral AI",
+      "weights": "209M weights",
+      "rel_crps": 0.446,
+      "crps_skill_pct": 55.4,
+      "rel_mase": 0.6575,
+      "mase_skill_pct": 34.25
+    },
+    {
+      "model_name": "Timestral-123M",
+      "provider": "Mistral AI",
+      "weights": "123M weights",
+      "rel_crps": 0.4476,
+      "crps_skill_pct": 55.24,
+      "rel_mase": 0.6589,
+      "mase_skill_pct": 34.11
+    },
+    {
+      "model_name": "TimesFM-3",
+      "provider": "Google",
+      "weights": "331M weights",
+      "rel_crps": 0.4557,
+      "crps_skill_pct": 54.43,
+      "rel_mase": 0.6668,
+      "mase_skill_pct": 33.32
+    },
+    {
+      "model_name": "Timestral-28M",
+      "provider": "Mistral AI",
+      "weights": "28M weights",
+      "rel_crps": 0.461,
+      "crps_skill_pct": 53.9,
+      "rel_mase": 0.678,
+      "mase_skill_pct": 32.2
+    },
+    {
+      "model_name": "Toto-2.0 (2.5B)",
+      "provider": "Datadog",
+      "weights": "2.5B weights",
+      "rel_crps": 0.4759,
+      "crps_skill_pct": 52.41,
+      "rel_mase": 0.6956,
+      "mase_skill_pct": 30.44
+    },
+    {
+      "model_name": "Chronos-2",
+      "provider": "AWS",
+      "weights": "120M weights",
+      "rel_crps": 0.4854,
+      "crps_skill_pct": 51.46,
+      "rel_mase": 0.6978,
+      "mase_skill_pct": 30.22
+    },
+    {
+      "model_name": "TimesFM-2.5",
+      "provider": "Google",
+      "weights": "200M weights",
+      "rel_crps": 0.4903,
+      "crps_skill_pct": 50.97,
+      "rel_mase": 0.705,
+      "mase_skill_pct": 29.5
+    },
+    {
+      "model_name": "Moirai-2.0",
+      "provider": "Salesforce",
+      "weights": "311M weights",
+      "rel_crps": 0.5164,
+      "crps_skill_pct": 48.36,
+      "rel_mase": 0.7281,
+      "mase_skill_pct": 27.19
+    },
+    {
+      "model_name": "TabPFN-TS",
+      "provider": "Prior Labs",
+      "weights": "88M weights",
+      "rel_crps": 0.5441,
+      "crps_skill_pct": 45.59,
+      "rel_mase": 0.7709,
+      "mase_skill_pct": 22.91
+    },
+    {
+      "model_name": "Telos-Small (50.1M Wts)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights (Standalone JEPA)",
+      "rel_crps": 0.5621,
+      "crps_skill_pct": 43.79,
+      "rel_mase": 0.7591,
+      "mase_skill_pct": 24.09
+    },
+    {
+      "model_name": "PatchTST",
+      "provider": "Transformer",
+      "weights": "1.8M weights",
+      "rel_crps": 0.587,
+      "crps_skill_pct": 41.3,
+      "rel_mase": 0.8486,
+      "mase_skill_pct": 15.14
+    },
+    {
+      "model_name": "TFT",
+      "provider": "Deep Learning",
+      "weights": "2.4M weights",
+      "rel_crps": 0.6047,
+      "crps_skill_pct": 39.53,
+      "rel_mase": 0.9153,
+      "mase_skill_pct": 8.47
+    },
+    {
+      "model_name": "DLinear",
+      "provider": "Linear Baseline",
+      "weights": "14K weights",
+      "rel_crps": 0.8456,
+      "crps_skill_pct": 15.44,
+      "rel_mase": 1.0608,
+      "mase_skill_pct": -6.08
+    },
+    {
+      "model_name": "Seasonal Naive",
+      "provider": "Baseline",
+      "weights": "0 weights",
+      "rel_crps": 1.0,
+      "crps_skill_pct": 0.0,
+      "rel_mase": 1.0,
+      "mase_skill_pct": 0.0
+    }
+  ],
+  "pure_nn_leaderboard": [
+    {
+      "model_name": "Telos-Small (Pure JEPA)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights",
+      "rel_crps": 0.5679,
+      "crps_skill_pct": 43.21,
+      "rel_mase": 0.8053,
+      "mase_skill_pct": 19.47
+    }
+  ],
+  "model_family": "Telos",
+  "model_name": "Telos-Small",
+  "parameters": 50100747,
+  "telos_small_parameters": 50100747
+}

results/ind_eval_telos_small_50m.json

@@ -0,0 +1,3905 @@
+{
+  "benchmark": "Cognite `IND-Eval` Industrial Benchmark (187 Tasks across 9 Physical/Industrial Sources \u2014 100% Standalone)",
+  "tasks_evaluated": 187,
+  "checkpoint_sha256": {
+    "telos_small": "ac7b7d2ff259aabdc7d94afccafae9790d2f9d3bb86097d16235ca10f7f2f7a7"
+  },
+  "parameters": 50100747,
+  "elapsed_s": 1906.7,
+  "leaderboard": [
+    {
+      "model_name": "Timestral-209M",
+      "provider": "Mistral AI",
+      "weights": "209M weights",
+      "ncrps": 0.4345,
+      "crps_skill_pct": 56.55,
+      "nmase": 0.646,
+      "mase_skill_pct": 35.4,
+      "core6_ncrps": 0.3911,
+      "core6_crps_skill_pct": 60.89,
+      "core6_nmase": 0.5911,
+      "per_source_ncrps": {
+        "3w": 0.5736,
+        "electricity": 0.5427,
+        "ett": 0.6357,
+        "mds": 0.4582,
+        "metropt3": 0.1902,
+        "neats": 0.4791,
+        "pronto": 0.4806,
+        "solar": 0.447,
+        "tablet_compression": 0.3109
+      }
+    },
+    {
+      "model_name": "Timestral-123M",
+      "provider": "Mistral AI",
+      "weights": "123M weights",
+      "ncrps": 0.4413,
+      "crps_skill_pct": 55.87,
+      "nmase": 0.6627,
+      "mase_skill_pct": 33.73,
+      "core6_ncrps": 0.3988,
+      "core6_crps_skill_pct": 60.12,
+      "core6_nmase": 0.6125,
+      "per_source_ncrps": {
+        "3w": 0.5894,
+        "electricity": 0.5474,
+        "ett": 0.6408,
+        "mds": 0.4613,
+        "metropt3": 0.1954,
+        "neats": 0.4775,
+        "pronto": 0.5122,
+        "solar": 0.4504,
+        "tablet_compression": 0.3094
+      }
+    },
+    {
+      "model_name": "Timestral-28M",
+      "provider": "Mistral AI",
+      "weights": "28M weights",
+      "ncrps": 0.4642,
+      "crps_skill_pct": 53.58,
+      "nmase": 0.6912,
+      "mase_skill_pct": 30.88,
+      "core6_ncrps": 0.4272,
+      "core6_crps_skill_pct": 57.28,
+      "core6_nmase": 0.6486,
+      "per_source_ncrps": {
+        "3w": 0.6169,
+        "electricity": 0.5566,
+        "ett": 0.643,
+        "mds": 0.4754,
+        "metropt3": 0.2175,
+        "neats": 0.4819,
+        "pronto": 0.6287,
+        "solar": 0.4602,
+        "tablet_compression": 0.3144
+      }
+    },
+    {
+      "model_name": "t0-gamma",
+      "provider": "The Forecasting Co.",
+      "weights": "256M weights",
+      "ncrps": 0.4675,
+      "crps_skill_pct": 53.25,
+      "nmase": 0.6941,
+      "mase_skill_pct": 30.59,
+      "core6_ncrps": 0.4522,
+      "core6_crps_skill_pct": 54.78,
+      "core6_nmase": 0.6832,
+      "per_source_ncrps": {
+        "3w": 0.6607,
+        "electricity": 0.5404,
+        "ett": 0.7017,
+        "mds": 0.5873,
+        "metropt3": 0.2534,
+        "neats": 0.5317,
+        "pronto": 0.5126,
+        "solar": 0.3289,
+        "tablet_compression": 0.3188
+      }
+    },
+    {
+      "model_name": "TimesFM-3",
+      "provider": "Google",
+      "weights": "331M weights",
+      "ncrps": 0.4734,
+      "crps_skill_pct": 52.66,
+      "nmase": 0.7127,
+      "mase_skill_pct": 28.73,
+      "core6_ncrps": 0.4357,
+      "core6_crps_skill_pct": 56.43,
+      "core6_nmase": 0.6735,
+      "per_source_ncrps": {
+        "3w": 0.69,
+        "electricity": 0.558,
+        "ett": 0.649,
+        "mds": 0.4814,
+        "metropt3": 0.2769,
+        "neats": 0.494,
+        "pronto": 0.4727,
+        "solar": 0.4819,
+        "tablet_compression": 0.3184
+      }
+    },
+    {
+      "model_name": "Timestral-5M",
+      "provider": "Mistral AI",
+      "weights": "5.2M weights",
+      "ncrps": 0.4876,
+      "crps_skill_pct": 51.24,
+      "nmase": 0.7256,
+      "mase_skill_pct": 27.44,
+      "core6_ncrps": 0.4516,
+      "core6_crps_skill_pct": 54.84,
+      "core6_nmase": 0.6861,
+      "per_source_ncrps": {
+        "3w": 0.6814,
+        "electricity": 0.5788,
+        "ett": 0.6544,
+        "mds": 0.4919,
+        "metropt3": 0.274,
+        "neats": 0.4842,
+        "pronto": 0.5909,
+        "solar": 0.4849,
+        "tablet_compression": 0.3229
+      }
+    },
+    {
+      "model_name": "Toto-2.0 (313M)",
+      "provider": "Datadog",
+      "weights": "313M weights",
+      "ncrps": 0.5011,
+      "crps_skill_pct": 49.89,
+      "nmase": 0.7349,
+      "mase_skill_pct": 26.51,
+      "core6_ncrps": 0.4705,
+      "core6_crps_skill_pct": 52.95,
+      "core6_nmase": 0.6997,
+      "per_source_ncrps": {
+        "3w": 0.6542,
+        "electricity": 0.5879,
+        "ett": 0.6655,
+        "mds": 0.4666,
+        "metropt3": 0.2614,
+        "neats": 0.4895,
+        "pronto": 0.8831,
+        "solar": 0.47,
+        "tablet_compression": 0.3143
+      }
+    },
+    {
+      "model_name": "t0-beta",
+      "provider": "The Forecasting Co.",
+      "weights": "256M weights",
+      "ncrps": 0.511,
+      "crps_skill_pct": 48.9,
+      "nmase": 0.7403,
+      "mase_skill_pct": 25.97,
+      "core6_ncrps": 0.5048,
+      "core6_crps_skill_pct": 49.52,
+      "core6_nmase": 0.7353,
+      "per_source_ncrps": {
+        "3w": 0.7156,
+        "electricity": 0.5735,
+        "ett": 0.6989,
+        "mds": 0.7346,
+        "metropt3": 0.2758,
+        "neats": 0.5723,
+        "pronto": 0.6123,
+        "solar": 0.3583,
+        "tablet_compression": 0.3256
+      }
+    },
+    {
+      "model_name": "Tafsut-base",
+      "provider": "Tafsut-FM",
+      "weights": "Foundation",
+      "ncrps": 0.5132,
+      "crps_skill_pct": 48.68,
+      "nmase": 0.7387,
+      "mase_skill_pct": 26.13,
+      "core6_ncrps": 0.4979,
+      "core6_crps_skill_pct": 50.21,
+      "core6_nmase": 0.7244,
+      "per_source_ncrps": {
+        "3w": 0.6796,
+        "electricity": 0.5641,
+        "ett": 0.6868,
+        "mds": 0.4911,
+        "metropt3": 0.3379,
+        "neats": 0.514,
+        "pronto": 0.8127,
+        "solar": 0.4183,
+        "tablet_compression": 0.3235
+      }
+    },
+    {
+      "model_name": "Granite-PatchTST-r2",
+      "provider": "IBM",
+      "weights": "Foundation",
+      "ncrps": 0.5209,
+      "crps_skill_pct": 47.91,
+      "nmase": 0.7661,
+      "mase_skill_pct": 23.39,
+      "core6_ncrps": 0.5053,
+      "core6_crps_skill_pct": 49.47,
+      "core6_nmase": 0.7517,
+      "per_source_ncrps": {
+        "3w": 0.6981,
+        "electricity": 0.5439,
+        "ett": 0.6683,
+        "mds": 0.4781,
+        "metropt3": 0.3302,
+        "neats": 0.4913,
+        "pronto": 0.9236,
+        "solar": 0.4663,
+        "tablet_compression": 0.3329
+      }
+    },
+    {
+      "model_name": "TiRex-2",
+      "provider": "NX-AI",
+      "weights": "82M weights",
+      "ncrps": 0.5368,
+      "crps_skill_pct": 46.32,
+      "nmase": 0.7792,
+      "mase_skill_pct": 22.08,
+      "core6_ncrps": 0.5088,
+      "core6_crps_skill_pct": 49.12,
+      "core6_nmase": 0.7567,
+      "per_source_ncrps": {
+        "3w": 0.7861,
+        "electricity": 0.5899,
+        "ett": 0.6916,
+        "mds": 0.5423,
+        "metropt3": 0.3214,
+        "neats": 0.5506,
+        "pronto": 0.7178,
+        "solar": 0.5226,
+        "tablet_compression": 0.3205
+      }
+    },
+    {
+      "model_name": "Chronos-2",
+      "provider": "AWS",
+      "weights": "120M weights",
+      "ncrps": 0.5386,
+      "crps_skill_pct": 46.14,
+      "nmase": 0.7547,
+      "mase_skill_pct": 24.53,
+      "core6_ncrps": 0.5282,
+      "core6_crps_skill_pct": 47.18,
+      "core6_nmase": 0.7318,
+      "per_source_ncrps": {
+        "3w": 0.7487,
+        "electricity": 0.5448,
+        "ett": 0.6739,
+        "mds": 0.9483,
+        "metropt3": 0.3216,
+        "neats": 0.5162,
+        "pronto": 0.5692,
+        "solar": 0.4784,
+        "tablet_compression": 0.3236
+      }
+    },
+    {
+      "model_name": "Telos-Small (50.1M Wts)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights (Standalone JEPA)",
+      "ncrps": 0.6594,
+      "crps_skill_pct": 34.06,
+      "nmase": 0.9411,
+      "mase_skill_pct": 5.89,
+      "core6_ncrps": 0.6794,
+      "core6_crps_skill_pct": 32.06,
+      "core6_nmase": 0.9898,
+      "per_source_ncrps": {
+        "3w": 1.0879,
+        "electricity": 0.6737,
+        "ett": 0.6787,
+        "mds": 0.7723,
+        "metropt3": 0.5261,
+        "neats": 0.5542,
+        "pronto": 1.1662,
+        "solar": 0.5243,
+        "tablet_compression": 0.3442
+      }
+    },
+    {
+      "model_name": "Seasonal Naive",
+      "provider": "Baseline",
+      "weights": "0 weights",
+      "ncrps": 1.0,
+      "crps_skill_pct": 0.0,
+      "nmase": 1.0,
+      "mase_skill_pct": 0.0,
+      "core6_ncrps": 1.0,
+      "core6_crps_skill_pct": 0.0,
+      "core6_nmase": 1.0,
+      "per_source_ncrps": {}
+    }
+  ],
+  "pure_nn_leaderboard": [
+    {
+      "model_name": "Telos-Small (Pure JEPA)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights",
+      "ncrps": 0.6182,
+      "crps_skill_pct": 38.18,
+      "nmase": 0.906,
+      "mase_skill_pct": 9.4,
+      "core6_ncrps": 0.6182,
+      "core6_crps_skill_pct": 38.18,
+      "core6_nmase": 0.93,
+      "per_source_ncrps": {
+        "3w": 1.0866,
+        "electricity": 0.6625,
+        "ett": 0.6806,
+        "mds": 0.6336,
+        "metropt3": 0.5224,
+        "neats": 0.5292,
+        "pronto": 0.8646,
+        "solar": 0.5244,
+        "tablet_compression": 0.3391
+      }
+    }
+  ],
+  "per_task": [
+    {
+      "source": "3w",
+      "dataset": "event-0",
+      "term": "long",
+      "H": 1024,
+      "items": 997,
+      "telos_small": {
+        "nCRPS": 0.7923,
+        "nMASE": 0.9613
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7585,
+        "nMASE": 0.951
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9522,
+        "nMASE": 0.9862
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-0",
+      "term": "short",
+      "H": 512,
+      "items": 1000,
+      "telos_small": {
+        "nCRPS": 0.8511,
+        "nMASE": 1.0298
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8034,
+        "nMASE": 1.0062
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0005,
+        "nMASE": 1.0546
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-1",
+      "term": "long",
+      "H": 1024,
+      "items": 680,
+      "telos_small": {
+        "nCRPS": 1.5408,
+        "nMASE": 1.2571
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.2956,
+        "nMASE": 1.2391
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.259,
+        "nMASE": 1.0112
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-1",
+      "term": "short",
+      "H": 512,
+      "items": 680,
+      "telos_small": {
+        "nCRPS": 1.4578,
+        "nMASE": 1.2977
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.2949,
+        "nMASE": 1.257
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.2415,
+        "nMASE": 1.0268
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-2",
+      "term": "long",
+      "H": 1024,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 1.3485,
+        "nMASE": 1.012
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 2.1653,
+        "nMASE": 1.3759
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.4331,
+        "nMASE": 1.006
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-2",
+      "term": "short",
+      "H": 512,
+      "items": 38,
+      "telos_small": {
+        "nCRPS": 1.5817,
+        "nMASE": 1.0257
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 2.7459,
+        "nMASE": 1.4979
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.9139,
+        "nMASE": 1.0183
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-3",
+      "term": "long",
+      "H": 1024,
+      "items": 1070,
+      "telos_small": {
+        "nCRPS": 0.6211,
+        "nMASE": 0.7462
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6027,
+        "nMASE": 0.7325
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8921,
+        "nMASE": 0.9113
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-3",
+      "term": "short",
+      "H": 512,
+      "items": 1070,
+      "telos_small": {
+        "nCRPS": 0.5743,
+        "nMASE": 0.7249
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5702,
+        "nMASE": 0.7218
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.92,
+        "nMASE": 0.9566
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-4",
+      "term": "long",
+      "H": 1024,
+      "items": 1380,
+      "telos_small": {
+        "nCRPS": 0.6029,
+        "nMASE": 0.8089
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.589,
+        "nMASE": 0.7881
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9098,
+        "nMASE": 0.9318
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-4",
+      "term": "short",
+      "H": 512,
+      "items": 1380,
+      "telos_small": {
+        "nCRPS": 0.54,
+        "nMASE": 0.7897
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5388,
+        "nMASE": 0.7853
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8781,
+        "nMASE": 0.9742
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-5",
+      "term": "long",
+      "H": 1024,
+      "items": 537,
+      "telos_small": {
+        "nCRPS": 0.8565,
+        "nMASE": 1.0225
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7998,
+        "nMASE": 1.1726
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8752,
+        "nMASE": 0.9873
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-5",
+      "term": "short",
+      "H": 512,
+      "items": 531,
+      "telos_small": {
+        "nCRPS": 0.8291,
+        "nMASE": 1.0091
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8879,
+        "nMASE": 1.1967
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8806,
+        "nMASE": 0.9745
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-6",
+      "term": "long",
+      "H": 1024,
+      "items": 23,
+      "telos_small": {
+        "nCRPS": 2.9833,
+        "nMASE": 1.5333
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 2.9408,
+        "nMASE": 2.5876
+      },
+      "analytical_baseline": {
+        "nCRPS": 3.0035,
+        "nMASE": 1.2577
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-6",
+      "term": "short",
+      "H": 512,
+      "items": 23,
+      "telos_small": {
+        "nCRPS": 1.3573,
+        "nMASE": 1.2107
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 2.4881,
+        "nMASE": 2.7219
+      },
+      "analytical_baseline": {
+        "nCRPS": 2.3701,
+        "nMASE": 1.204
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-7",
+      "term": "long",
+      "H": 1024,
+      "items": 100,
+      "telos_small": {
+        "nCRPS": 1.0913,
+        "nMASE": 1.1491
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.0281,
+        "nMASE": 1.1287
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.303,
+        "nMASE": 1.0105
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-7",
+      "term": "short",
+      "H": 512,
+      "items": 100,
+      "telos_small": {
+        "nCRPS": 2.3379,
+        "nMASE": 1.8048
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.1748,
+        "nMASE": 1.3218
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.2746,
+        "nMASE": 1.0387
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-8",
+      "term": "long",
+      "H": 1024,
+      "items": 119,
+      "telos_small": {
+        "nCRPS": 1.1153,
+        "nMASE": 1.3831
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8784,
+        "nMASE": 0.9238
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.2054,
+        "nMASE": 1.0613
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-8",
+      "term": "short",
+      "H": 512,
+      "items": 119,
+      "telos_small": {
+        "nCRPS": 1.1406,
+        "nMASE": 1.1598
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.9175,
+        "nMASE": 1.0219
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.2516,
+        "nMASE": 1.1685
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-9",
+      "term": "long",
+      "H": 1024,
+      "items": 776,
+      "telos_small": {
+        "nCRPS": 1.331,
+        "nMASE": 1.4237
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.3579,
+        "nMASE": 1.5959
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.5505,
+        "nMASE": 1.074
+      }
+    },
+    {
+      "source": "3w",
+      "dataset": "event-9",
+      "term": "short",
+      "H": 512,
+      "items": 776,
+      "telos_small": {
+        "nCRPS": 1.1571,
+        "nMASE": 1.1939
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.1568,
+        "nMASE": 1.3751
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.3816,
+        "nMASE": 1.0795
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-15T",
+      "term": "long",
+      "H": 720,
+      "items": 7398,
+      "telos_small": {
+        "nCRPS": 0.6984,
+        "nMASE": 0.7936
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7438,
+        "nMASE": 0.8449
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8287,
+        "nMASE": 0.8763
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-15T",
+      "term": "medium",
+      "H": 480,
+      "items": 7400,
+      "telos_small": {
+        "nCRPS": 0.7467,
+        "nMASE": 0.8077
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7712,
+        "nMASE": 0.8329
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.837,
+        "nMASE": 0.8786
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-15T",
+      "term": "short",
+      "H": 48,
+      "items": 7400,
+      "telos_small": {
+        "nCRPS": 0.6504,
+        "nMASE": 0.7048
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6409,
+        "nMASE": 0.7064
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7478,
+        "nMASE": 0.7778
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-D",
+      "term": "short",
+      "H": 30,
+      "items": 1697,
+      "telos_small": {
+        "nCRPS": 0.6732,
+        "nMASE": 0.7943
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6693,
+        "nMASE": 0.792
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8236,
+        "nMASE": 0.8878
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-H",
+      "term": "long",
+      "H": 720,
+      "items": 1846,
+      "telos_small": {
+        "nCRPS": 0.619,
+        "nMASE": 0.8128
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5853,
+        "nMASE": 0.8125
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7956,
+        "nMASE": 0.8837
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-H",
+      "term": "medium",
+      "H": 480,
+      "items": 2957,
+      "telos_small": {
+        "nCRPS": 0.6424,
+        "nMASE": 0.801
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6063,
+        "nMASE": 0.8029
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8116,
+        "nMASE": 0.8684
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-H",
+      "term": "short",
+      "H": 48,
+      "items": 7400,
+      "telos_small": {
+        "nCRPS": 0.6787,
+        "nMASE": 0.7376
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6897,
+        "nMASE": 0.7519
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7888,
+        "nMASE": 0.8183
+      }
+    },
+    {
+      "source": "electricity",
+      "dataset": "electricity-W",
+      "term": "short",
+      "H": 8,
+      "items": 1019,
+      "telos_small": {
+        "nCRPS": 0.6887,
+        "nMASE": 0.839
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6162,
+        "nMASE": 0.8136
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.1304,
+        "nMASE": 1.0156
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-15T",
+      "term": "long",
+      "H": 720,
+      "items": 70,
+      "telos_small": {
+        "nCRPS": 0.6789,
+        "nMASE": 0.8703
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6716,
+        "nMASE": 0.8874
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7551,
+        "nMASE": 0.8943
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-15T",
+      "term": "medium",
+      "H": 480,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.7296,
+        "nMASE": 0.8539
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7339,
+        "nMASE": 0.8706
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7936,
+        "nMASE": 0.8615
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-15T",
+      "term": "short",
+      "H": 48,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.6398,
+        "nMASE": 0.7404
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6401,
+        "nMASE": 0.7424
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7625,
+        "nMASE": 0.7882
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-D",
+      "term": "short",
+      "H": 30,
+      "items": 21,
+      "telos_small": {
+        "nCRPS": 0.6861,
+        "nMASE": 0.9796
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6826,
+        "nMASE": 0.9868
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8354,
+        "nMASE": 0.9891
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-H",
+      "term": "long",
+      "H": 720,
+      "items": 21,
+      "telos_small": {
+        "nCRPS": 0.55,
+        "nMASE": 0.8838
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5367,
+        "nMASE": 0.886
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6631,
+        "nMASE": 0.9109
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-H",
+      "term": "medium",
+      "H": 480,
+      "items": 28,
+      "telos_small": {
+        "nCRPS": 0.5679,
+        "nMASE": 0.7981
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5661,
+        "nMASE": 0.7971
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6849,
+        "nMASE": 0.8649
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-H",
+      "term": "short",
+      "H": 48,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.7542,
+        "nMASE": 0.8387
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.755,
+        "nMASE": 0.8392
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8493,
+        "nMASE": 0.9056
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett1-W",
+      "term": "short",
+      "H": 8,
+      "items": 14,
+      "telos_small": {
+        "nCRPS": 0.799,
+        "nMASE": 0.8188
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.842,
+        "nMASE": 0.8968
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8548,
+        "nMASE": 0.8339
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-15T",
+      "term": "long",
+      "H": 720,
+      "items": 70,
+      "telos_small": {
+        "nCRPS": 0.7663,
+        "nMASE": 0.8913
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7683,
+        "nMASE": 0.9035
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7948,
+        "nMASE": 0.9093
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-15T",
+      "term": "medium",
+      "H": 480,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.8065,
+        "nMASE": 0.8422
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8264,
+        "nMASE": 0.8679
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8328,
+        "nMASE": 0.86
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-15T",
+      "term": "short",
+      "H": 48,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.6945,
+        "nMASE": 0.727
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6973,
+        "nMASE": 0.7307
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.775,
+        "nMASE": 0.773
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-D",
+      "term": "short",
+      "H": 30,
+      "items": 21,
+      "telos_small": {
+        "nCRPS": 0.6407,
+        "nMASE": 1.0371
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6489,
+        "nMASE": 1.0602
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8438,
+        "nMASE": 1.0443
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-H",
+      "term": "long",
+      "H": 720,
+      "items": 21,
+      "telos_small": {
+        "nCRPS": 0.5789,
+        "nMASE": 0.9617
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5595,
+        "nMASE": 0.9549
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6194,
+        "nMASE": 0.9243
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-H",
+      "term": "medium",
+      "H": 480,
+      "items": 28,
+      "telos_small": {
+        "nCRPS": 0.5991,
+        "nMASE": 0.8504
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6141,
+        "nMASE": 0.8763
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7232,
+        "nMASE": 0.9466
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-H",
+      "term": "short",
+      "H": 48,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.7773,
+        "nMASE": 0.8292
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7803,
+        "nMASE": 0.8333
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.835,
+        "nMASE": 0.8727
+      }
+    },
+    {
+      "source": "ett",
+      "dataset": "ett2-W",
+      "term": "short",
+      "H": 8,
+      "items": 14,
+      "telos_small": {
+        "nCRPS": 0.6684,
+        "nMASE": 1.0007
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.661,
+        "nMASE": 0.977
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9424,
+        "nMASE": 1.6219
+      }
+    },
+    {
+      "source": "mds",
+      "dataset": "mds-10s",
+      "term": "short",
+      "H": 512,
+      "items": 107430,
+      "telos_small": {
+        "nCRPS": 1.067,
+        "nMASE": 1.5556
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5853,
+        "nMASE": 1.2083
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8294,
+        "nMASE": 1.0267
+      }
+    },
+    {
+      "source": "mds",
+      "dataset": "mds-2min",
+      "term": "short",
+      "H": 512,
+      "items": 6495,
+      "telos_small": {
+        "nCRPS": 0.726,
+        "nMASE": 1.0473
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7271,
+        "nMASE": 1.1462
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7707,
+        "nMASE": 0.907
+      }
+    },
+    {
+      "source": "mds",
+      "dataset": "mds-5min",
+      "term": "short",
+      "H": 512,
+      "items": 1590,
+      "telos_small": {
+        "nCRPS": 0.5946,
+        "nMASE": 0.9433
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5976,
+        "nMASE": 0.9364
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6219,
+        "nMASE": 0.9248
+      }
+    },
+    {
+      "source": "metropt3",
+      "dataset": "metropt3",
+      "term": "long",
+      "H": 180,
+      "items": 950,
+      "telos_small": {
+        "nCRPS": 0.4874,
+        "nMASE": 0.6885
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4799,
+        "nMASE": 0.6749
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6956,
+        "nMASE": 0.929
+      }
+    },
+    {
+      "source": "metropt3",
+      "dataset": "metropt3",
+      "term": "medium",
+      "H": 60,
+      "items": 950,
+      "telos_small": {
+        "nCRPS": 0.5117,
+        "nMASE": 0.6241
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5111,
+        "nMASE": 0.6232
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7782,
+        "nMASE": 0.9298
+      }
+    },
+    {
+      "source": "metropt3",
+      "dataset": "metropt3",
+      "term": "short",
+      "H": 30,
+      "items": 950,
+      "telos_small": {
+        "nCRPS": 0.5837,
+        "nMASE": 0.7228
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5812,
+        "nMASE": 0.7199
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8377,
+        "nMASE": 1.0099
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "banambu-deep-1__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 151,
+      "telos_small": {
+        "nCRPS": 0.5575,
+        "nMASE": 1.0079
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5588,
+        "nMASE": 1.0381
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8308,
+        "nMASE": 1.0408
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "banambu-deep-1__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 162,
+      "telos_small": {
+        "nCRPS": 0.6167,
+        "nMASE": 1.0414
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6104,
+        "nMASE": 1.0094
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8702,
+        "nMASE": 1.1096
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bassett-deep-1__mwd-lwd-dml",
+      "term": "long",
+      "H": 1024,
+      "items": 218,
+      "telos_small": {
+        "nCRPS": 0.7032,
+        "nMASE": 1.1115
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6686,
+        "nMASE": 1.058
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9493,
+        "nMASE": 1.0195
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bassett-deep-1__mwd-lwd-dml",
+      "term": "short",
+      "H": 512,
+      "items": 229,
+      "telos_small": {
+        "nCRPS": 0.7163,
+        "nMASE": 1.1193
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6821,
+        "nMASE": 1.1005
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0078,
+        "nMASE": 1.0471
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-01__mwd-lwd-dml-st1",
+      "term": "long",
+      "H": 1024,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.6781,
+        "nMASE": 1.0726
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7205,
+        "nMASE": 1.1334
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8523,
+        "nMASE": 1.0189
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-01__mwd-lwd-dml-st1",
+      "term": "short",
+      "H": 512,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.7315,
+        "nMASE": 1.1864
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7393,
+        "nMASE": 1.196
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8813,
+        "nMASE": 1.0807
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-01__mwd-lwd-dml",
+      "term": "long",
+      "H": 1024,
+      "items": 72,
+      "telos_small": {
+        "nCRPS": 0.6043,
+        "nMASE": 0.9229
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6198,
+        "nMASE": 0.9438
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9054,
+        "nMASE": 1.0155
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-01__mwd-lwd-dml",
+      "term": "short",
+      "H": 512,
+      "items": 80,
+      "telos_small": {
+        "nCRPS": 0.6711,
+        "nMASE": 0.9769
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6722,
+        "nMASE": 0.978
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.874,
+        "nMASE": 0.9907
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-04-phase-2__mwd-lwd-dml",
+      "term": "long",
+      "H": 1024,
+      "items": 95,
+      "telos_small": {
+        "nCRPS": 0.9332,
+        "nMASE": 1.2419
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8393,
+        "nMASE": 1.311
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0575,
+        "nMASE": 1.1326
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-2-04-phase-2__mwd-lwd-dml",
+      "term": "short",
+      "H": 512,
+      "items": 95,
+      "telos_small": {
+        "nCRPS": 0.8705,
+        "nMASE": 1.2244
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8741,
+        "nMASE": 1.2744
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0641,
+        "nMASE": 1.1337
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-3-05-p__mwd-lwd-dml",
+      "term": "long",
+      "H": 1024,
+      "items": 159,
+      "telos_small": {
+        "nCRPS": 0.7443,
+        "nMASE": 0.999
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7916,
+        "nMASE": 1.1092
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9244,
+        "nMASE": 0.9983
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "bdc-3-05-p__mwd-lwd-dml",
+      "term": "short",
+      "H": 512,
+      "items": 159,
+      "telos_small": {
+        "nCRPS": 0.7508,
+        "nMASE": 0.9863
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.8201,
+        "nMASE": 1.1349
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9456,
+        "nMASE": 1.0247
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "gemtree-1__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 109,
+      "telos_small": {
+        "nCRPS": 0.5297,
+        "nMASE": 1.1851
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4902,
+        "nMASE": 1.1222
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.704,
+        "nMASE": 1.0705
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "gemtree-1__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 109,
+      "telos_small": {
+        "nCRPS": 0.5106,
+        "nMASE": 0.8605
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5439,
+        "nMASE": 0.9739
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7302,
+        "nMASE": 0.9382
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "holonema-1__mwd-lwd-dml",
+      "term": "long",
+      "H": 1024,
+      "items": 205,
+      "telos_small": {
+        "nCRPS": 0.4317,
+        "nMASE": 2.8964
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4814,
+        "nMASE": 3.119
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7011,
+        "nMASE": 2.8681
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "holonema-1__mwd-lwd-dml",
+      "term": "short",
+      "H": 512,
+      "items": 205,
+      "telos_small": {
+        "nCRPS": 0.454,
+        "nMASE": 3.591
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4929,
+        "nMASE": 3.906
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7403,
+        "nMASE": 3.6764
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "la-rocca-1__lwd-dds",
+      "term": "long",
+      "H": 1024,
+      "items": 68,
+      "telos_small": {
+        "nCRPS": 0.2722,
+        "nMASE": 0.7606
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2681,
+        "nMASE": 0.7647
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4681,
+        "nMASE": 1.0051
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "la-rocca-1__lwd-dds",
+      "term": "short",
+      "H": 512,
+      "items": 76,
+      "telos_small": {
+        "nCRPS": 0.3503,
+        "nMASE": 0.7348
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3491,
+        "nMASE": 0.7378
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.556,
+        "nMASE": 0.9948
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "la-rocca-1__lwd-pwd",
+      "term": "long",
+      "H": 1024,
+      "items": 20,
+      "telos_small": {
+        "nCRPS": 0.1482,
+        "nMASE": 0.9893
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1453,
+        "nMASE": 0.9917
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.2668,
+        "nMASE": 1.0075
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "la-rocca-1__lwd-pwd",
+      "term": "short",
+      "H": 512,
+      "items": 20,
+      "telos_small": {
+        "nCRPS": 0.2116,
+        "nMASE": 0.9852
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2127,
+        "nMASE": 0.9863
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.3158,
+        "nMASE": 0.9959
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mudlogging",
+      "term": "long",
+      "H": 1024,
+      "items": 775,
+      "telos_small": {
+        "nCRPS": 0.6661,
+        "nMASE": 1.0501
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6609,
+        "nMASE": 1.0574
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0831,
+        "nMASE": 1.0278
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mudlogging",
+      "term": "short",
+      "H": 512,
+      "items": 775,
+      "telos_small": {
+        "nCRPS": 0.806,
+        "nMASE": 1.3681
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7273,
+        "nMASE": 1.0844
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.1618,
+        "nMASE": 1.0635
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 63,
+      "telos_small": {
+        "nCRPS": 0.6884,
+        "nMASE": 1.0868
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5272,
+        "nMASE": 0.8534
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8133,
+        "nMASE": 1.0017
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 67,
+      "telos_small": {
+        "nCRPS": 0.8205,
+        "nMASE": 1.1499
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6207,
+        "nMASE": 0.9316
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8636,
+        "nMASE": 1.0254
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mwd-lwd",
+      "term": "long",
+      "H": 1024,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.7696,
+        "nMASE": 1.0079
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.668,
+        "nMASE": 1.0455
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.948,
+        "nMASE": 1.0535
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-01__mwd-lwd",
+      "term": "short",
+      "H": 512,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.6559,
+        "nMASE": 1.0083
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6446,
+        "nMASE": 1.0127
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9529,
+        "nMASE": 1.1686
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-02__mudlogging",
+      "term": "long",
+      "H": 1024,
+      "items": 1115,
+      "telos_small": {
+        "nCRPS": 0.9194,
+        "nMASE": 1.3994
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6288,
+        "nMASE": 0.9115
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9831,
+        "nMASE": 0.994
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-02__mudlogging",
+      "term": "short",
+      "H": 512,
+      "items": 1115,
+      "telos_small": {
+        "nCRPS": 0.8659,
+        "nMASE": 1.2711
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7134,
+        "nMASE": 0.9733
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0929,
+        "nMASE": 1.0108
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-02__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 90,
+      "telos_small": {
+        "nCRPS": 0.1494,
+        "nMASE": 1.0339
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1482,
+        "nMASE": 1.0538
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.2249,
+        "nMASE": 0.9865
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-02__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 90,
+      "telos_small": {
+        "nCRPS": 0.1617,
+        "nMASE": 0.9435
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1609,
+        "nMASE": 0.9617
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.2465,
+        "nMASE": 0.9959
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-03__mudlogging",
+      "term": "long",
+      "H": 1024,
+      "items": 1029,
+      "telos_small": {
+        "nCRPS": 0.7419,
+        "nMASE": 1.23
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7339,
+        "nMASE": 1.1094
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0811,
+        "nMASE": 1.0376
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-03__mudlogging",
+      "term": "short",
+      "H": 512,
+      "items": 1031,
+      "telos_small": {
+        "nCRPS": 0.9024,
+        "nMASE": 1.2172
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.754,
+        "nMASE": 1.1377
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.1862,
+        "nMASE": 1.0507
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-03__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 115,
+      "telos_small": {
+        "nCRPS": 0.8196,
+        "nMASE": 2.1719
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7134,
+        "nMASE": 2.1243
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9421,
+        "nMASE": 2.2399
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "lpa-03__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 115,
+      "telos_small": {
+        "nCRPS": 0.7093,
+        "nMASE": 2.6943
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6998,
+        "nMASE": 2.6517
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9591,
+        "nMASE": 2.8455
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mudlogging-gravel-pack",
+      "term": "long",
+      "H": 1024,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.1823,
+        "nMASE": 0.5782
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1817,
+        "nMASE": 0.5722
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4562,
+        "nMASE": 0.8609
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mudlogging-gravel-pack",
+      "term": "short",
+      "H": 512,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.2226,
+        "nMASE": 0.5728
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2195,
+        "nMASE": 0.5623
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5318,
+        "nMASE": 0.8791
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mudlogging",
+      "term": "long",
+      "H": 1024,
+      "items": 461,
+      "telos_small": {
+        "nCRPS": 0.8074,
+        "nMASE": 1.2276
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6723,
+        "nMASE": 1.0151
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9579,
+        "nMASE": 1.046
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mudlogging",
+      "term": "short",
+      "H": 512,
+      "items": 461,
+      "telos_small": {
+        "nCRPS": 0.7603,
+        "nMASE": 3.4402
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6892,
+        "nMASE": 1.0013
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0496,
+        "nMASE": 1.0462
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mwd-lwd-drilling-mechanics",
+      "term": "long",
+      "H": 1024,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.6216,
+        "nMASE": 3.9579
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6085,
+        "nMASE": 3.95
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.788,
+        "nMASE": 4.0934
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "psa-02-phase-1__mwd-lwd-drilling-mechanics",
+      "term": "short",
+      "H": 512,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.7688,
+        "nMASE": 3.5366
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7261,
+        "nMASE": 3.5194
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.82,
+        "nMASE": 3.5851
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "sca-03__lwd-pwd",
+      "term": "long",
+      "H": 1024,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.6669,
+        "nMASE": 0.9757
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6688,
+        "nMASE": 0.976
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8442,
+        "nMASE": 1.0271
+      }
+    },
+    {
+      "source": "neats",
+      "dataset": "sca-03__lwd-pwd",
+      "term": "short",
+      "H": 512,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.7495,
+        "nMASE": 1.0166
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7503,
+        "nMASE": 1.0175
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9166,
+        "nMASE": 1.0391
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D2E1",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 1.0069,
+        "nMASE": 1.0251
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.9622,
+        "nMASE": 0.9901
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9788,
+        "nMASE": 1.0021
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D2E1",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.9719,
+        "nMASE": 0.9999
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.9956,
+        "nMASE": 0.9989
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9848,
+        "nMASE": 1.0019
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D2E2",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.642,
+        "nMASE": 0.732
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6077,
+        "nMASE": 0.6502
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7523,
+        "nMASE": 0.9444
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D2E2",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.8281,
+        "nMASE": 0.9558
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.775,
+        "nMASE": 0.7802
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8242,
+        "nMASE": 0.9562
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D3E1",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 2.2792,
+        "nMASE": 1.1316
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.0036,
+        "nMASE": 0.9864
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.2524,
+        "nMASE": 1.0152
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D3E1",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 6.6725,
+        "nMASE": 1.6593
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.9649,
+        "nMASE": 0.9886
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.214,
+        "nMASE": 1.015
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D3E2",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 1.0794,
+        "nMASE": 0.84
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.0795,
+        "nMASE": 0.84
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.1024,
+        "nMASE": 0.897
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D3E2",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 1.0406,
+        "nMASE": 0.886
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 1.0529,
+        "nMASE": 0.8899
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0689,
+        "nMASE": 0.9181
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D4E1",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.8972,
+        "nMASE": 0.9271
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7844,
+        "nMASE": 0.645
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9153,
+        "nMASE": 0.9661
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D4E1",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.9541,
+        "nMASE": 0.9371
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.913,
+        "nMASE": 0.8212
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9226,
+        "nMASE": 0.9301
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D4E2",
+      "term": "long",
+      "H": 1024,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.845,
+        "nMASE": 1.0176
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6856,
+        "nMASE": 0.6269
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8578,
+        "nMASE": 1.0191
+      }
+    },
+    {
+      "source": "pronto",
+      "dataset": "pronto_D4E2",
+      "term": "short",
+      "H": 512,
+      "items": 12,
+      "telos_small": {
+        "nCRPS": 0.9843,
+        "nMASE": 1.1253
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7153,
+        "nMASE": 0.7786
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8754,
+        "nMASE": 1.0192
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-10T",
+      "term": "long",
+      "H": 720,
+      "items": 1096,
+      "telos_small": {
+        "nCRPS": 0.4773,
+        "nMASE": 0.9376
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4759,
+        "nMASE": 0.94
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6048,
+        "nMASE": 0.9891
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-10T",
+      "term": "medium",
+      "H": 480,
+      "items": 1507,
+      "telos_small": {
+        "nCRPS": 0.5249,
+        "nMASE": 0.8587
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5199,
+        "nMASE": 0.8521
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.699,
+        "nMASE": 0.9875
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-10T",
+      "term": "short",
+      "H": 48,
+      "items": 2740,
+      "telos_small": {
+        "nCRPS": 0.7181,
+        "nMASE": 1.0074
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7094,
+        "nMASE": 0.9939
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8934,
+        "nMASE": 1.09
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-D",
+      "term": "short",
+      "H": 30,
+      "items": 274,
+      "telos_small": {
+        "nCRPS": 0.5201,
+        "nMASE": 0.8519
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.531,
+        "nMASE": 0.8599
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6433,
+        "nMASE": 0.8827
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-H",
+      "term": "long",
+      "H": 720,
+      "items": 274,
+      "telos_small": {
+        "nCRPS": 0.3177,
+        "nMASE": 0.8682
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3276,
+        "nMASE": 0.9141
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.387,
+        "nMASE": 0.9231
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-H",
+      "term": "medium",
+      "H": 480,
+      "items": 274,
+      "telos_small": {
+        "nCRPS": 0.3998,
+        "nMASE": 0.9475
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3988,
+        "nMASE": 0.9754
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4569,
+        "nMASE": 0.9537
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-H",
+      "term": "short",
+      "H": 48,
+      "items": 2603,
+      "telos_small": {
+        "nCRPS": 0.6101,
+        "nMASE": 0.9353
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6024,
+        "nMASE": 0.9399
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6857,
+        "nMASE": 0.8926
+      }
+    },
+    {
+      "source": "solar",
+      "dataset": "solar-W",
+      "term": "short",
+      "H": 8,
+      "items": 137,
+      "telos_small": {
+        "nCRPS": 0.7871,
+        "nMASE": 0.8465
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.7792,
+        "nMASE": 0.8383
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.946,
+        "nMASE": 0.99
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_01",
+      "term": "long",
+      "H": 180,
+      "items": 340,
+      "telos_small": {
+        "nCRPS": 0.269,
+        "nMASE": 0.792
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2685,
+        "nMASE": 0.789
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6819,
+        "nMASE": 0.8843
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_01",
+      "term": "medium",
+      "H": 60,
+      "items": 340,
+      "telos_small": {
+        "nCRPS": 0.3982,
+        "nMASE": 0.8417
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3876,
+        "nMASE": 0.7619
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7536,
+        "nMASE": 0.8769
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_01",
+      "term": "short",
+      "H": 30,
+      "items": 340,
+      "telos_small": {
+        "nCRPS": 0.487,
+        "nMASE": 0.7859
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4859,
+        "nMASE": 0.766
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8035,
+        "nMASE": 0.8794
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_02",
+      "term": "long",
+      "H": 180,
+      "items": 80,
+      "telos_small": {
+        "nCRPS": 0.2319,
+        "nMASE": 0.7695
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2327,
+        "nMASE": 0.7698
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5734,
+        "nMASE": 0.8418
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_02",
+      "term": "medium",
+      "H": 60,
+      "items": 80,
+      "telos_small": {
+        "nCRPS": 0.3425,
+        "nMASE": 0.7533
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3429,
+        "nMASE": 0.7523
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6363,
+        "nMASE": 0.8173
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_02",
+      "term": "short",
+      "H": 30,
+      "items": 80,
+      "telos_small": {
+        "nCRPS": 0.4243,
+        "nMASE": 0.7595
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4244,
+        "nMASE": 0.7587
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6713,
+        "nMASE": 0.8111
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_03",
+      "term": "long",
+      "H": 180,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.2215,
+        "nMASE": 0.6129
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2146,
+        "nMASE": 0.5174
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6319,
+        "nMASE": 0.9749
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_03",
+      "term": "medium",
+      "H": 60,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.3034,
+        "nMASE": 0.5207
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3017,
+        "nMASE": 0.5034
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7051,
+        "nMASE": 0.9544
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_03",
+      "term": "short",
+      "H": 30,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.3931,
+        "nMASE": 0.5424
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3908,
+        "nMASE": 0.5267
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7518,
+        "nMASE": 0.946
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_04",
+      "term": "long",
+      "H": 180,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.2057,
+        "nMASE": 1.2541
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2056,
+        "nMASE": 0.8212
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5881,
+        "nMASE": 0.9249
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_04",
+      "term": "medium",
+      "H": 60,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.3105,
+        "nMASE": 0.8014
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.311,
+        "nMASE": 0.8037
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6602,
+        "nMASE": 0.921
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_04",
+      "term": "short",
+      "H": 30,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.3948,
+        "nMASE": 0.7974
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3972,
+        "nMASE": 0.807
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7033,
+        "nMASE": 0.9196
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_05",
+      "term": "long",
+      "H": 180,
+      "items": 15,
+      "telos_small": {
+        "nCRPS": 0.4633,
+        "nMASE": 0.6389
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.46,
+        "nMASE": 0.6272
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8353,
+        "nMASE": 0.8059
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_05",
+      "term": "medium",
+      "H": 60,
+      "items": 15,
+      "telos_small": {
+        "nCRPS": 0.4158,
+        "nMASE": 0.5673
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4155,
+        "nMASE": 0.567
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.9523,
+        "nMASE": 0.7534
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_05",
+      "term": "short",
+      "H": 30,
+      "items": 15,
+      "telos_small": {
+        "nCRPS": 0.5108,
+        "nMASE": 0.5911
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5178,
+        "nMASE": 0.5823
+      },
+      "analytical_baseline": {
+        "nCRPS": 1.0018,
+        "nMASE": 0.7601
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_06",
+      "term": "long",
+      "H": 180,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.1944,
+        "nMASE": 0.8393
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1943,
+        "nMASE": 0.8392
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6424,
+        "nMASE": 0.8835
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_06",
+      "term": "medium",
+      "H": 60,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.2756,
+        "nMASE": 0.7585
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.275,
+        "nMASE": 0.7583
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7091,
+        "nMASE": 0.7989
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_06",
+      "term": "short",
+      "H": 30,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.3735,
+        "nMASE": 0.745
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3743,
+        "nMASE": 0.7459
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7555,
+        "nMASE": 0.8095
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_07",
+      "term": "long",
+      "H": 180,
+      "items": 35,
+      "telos_small": {
+        "nCRPS": 0.2178,
+        "nMASE": 0.6517
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2082,
+        "nMASE": 0.5333
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5814,
+        "nMASE": 0.8957
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_07",
+      "term": "medium",
+      "H": 60,
+      "items": 35,
+      "telos_small": {
+        "nCRPS": 0.2952,
+        "nMASE": 0.5371
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2978,
+        "nMASE": 0.5174
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6611,
+        "nMASE": 0.902
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_07",
+      "term": "short",
+      "H": 30,
+      "items": 35,
+      "telos_small": {
+        "nCRPS": 0.3776,
+        "nMASE": 0.4923
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3767,
+        "nMASE": 0.4745
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7176,
+        "nMASE": 0.8819
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_08",
+      "term": "long",
+      "H": 180,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.2356,
+        "nMASE": 0.8517
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2357,
+        "nMASE": 0.8133
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6906,
+        "nMASE": 0.9072
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_08",
+      "term": "medium",
+      "H": 60,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.3398,
+        "nMASE": 0.8125
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.34,
+        "nMASE": 0.7824
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8003,
+        "nMASE": 0.9297
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_08",
+      "term": "short",
+      "H": 30,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.4248,
+        "nMASE": 0.7602
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4254,
+        "nMASE": 0.7594
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8583,
+        "nMASE": 0.9395
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_09",
+      "term": "long",
+      "H": 180,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.3136,
+        "nMASE": 0.9122
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3134,
+        "nMASE": 0.8275
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6718,
+        "nMASE": 0.9575
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_09",
+      "term": "medium",
+      "H": 60,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.4987,
+        "nMASE": 5.2914
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4961,
+        "nMASE": 0.7886
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7503,
+        "nMASE": 0.9459
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_09",
+      "term": "short",
+      "H": 30,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.6202,
+        "nMASE": 2.5335
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6139,
+        "nMASE": 0.7875
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8076,
+        "nMASE": 0.8897
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_10",
+      "term": "long",
+      "H": 180,
+      "items": 135,
+      "telos_small": {
+        "nCRPS": 0.3672,
+        "nMASE": 0.8475
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3628,
+        "nMASE": 0.7708
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6455,
+        "nMASE": 0.8867
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_10",
+      "term": "medium",
+      "H": 60,
+      "items": 135,
+      "telos_small": {
+        "nCRPS": 0.424,
+        "nMASE": 0.8844
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4172,
+        "nMASE": 0.7468
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7056,
+        "nMASE": 0.8758
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_10",
+      "term": "short",
+      "H": 30,
+      "items": 135,
+      "telos_small": {
+        "nCRPS": 0.4366,
+        "nMASE": 0.7839
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4323,
+        "nMASE": 0.7261
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7012,
+        "nMASE": 0.8494
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_11",
+      "term": "long",
+      "H": 180,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.194,
+        "nMASE": 0.9789
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1944,
+        "nMASE": 0.8627
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.3482,
+        "nMASE": 1.0543
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_11",
+      "term": "medium",
+      "H": 60,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.2697,
+        "nMASE": 0.7777
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2702,
+        "nMASE": 0.7399
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4326,
+        "nMASE": 0.9433
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_11",
+      "term": "short",
+      "H": 30,
+      "items": 65,
+      "telos_small": {
+        "nCRPS": 0.3067,
+        "nMASE": 0.8093
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3078,
+        "nMASE": 0.8013
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4911,
+        "nMASE": 0.9292
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_12",
+      "term": "long",
+      "H": 180,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.25,
+        "nMASE": 1.1838
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2494,
+        "nMASE": 0.8591
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6621,
+        "nMASE": 0.9116
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_12",
+      "term": "medium",
+      "H": 60,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.3605,
+        "nMASE": 1.0095
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3599,
+        "nMASE": 0.815
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7244,
+        "nMASE": 0.8966
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_12",
+      "term": "short",
+      "H": 30,
+      "items": 105,
+      "telos_small": {
+        "nCRPS": 0.4725,
+        "nMASE": 0.886
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4726,
+        "nMASE": 0.8399
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7655,
+        "nMASE": 0.8908
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_13",
+      "term": "long",
+      "H": 180,
+      "items": 655,
+      "telos_small": {
+        "nCRPS": 0.2112,
+        "nMASE": 1.5075
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1852,
+        "nMASE": 0.7684
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4462,
+        "nMASE": 0.9303
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_13",
+      "term": "medium",
+      "H": 60,
+      "items": 655,
+      "telos_small": {
+        "nCRPS": 0.3016,
+        "nMASE": 0.9833
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2815,
+        "nMASE": 0.7802
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.507,
+        "nMASE": 0.928
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_13",
+      "term": "short",
+      "H": 30,
+      "items": 655,
+      "telos_small": {
+        "nCRPS": 0.3699,
+        "nMASE": 0.8645
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3612,
+        "nMASE": 0.8335
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5568,
+        "nMASE": 0.9351
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_14",
+      "term": "long",
+      "H": 180,
+      "items": 55,
+      "telos_small": {
+        "nCRPS": 0.2095,
+        "nMASE": 0.9672
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2101,
+        "nMASE": 0.9649
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6013,
+        "nMASE": 0.9644
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_14",
+      "term": "medium",
+      "H": 60,
+      "items": 55,
+      "telos_small": {
+        "nCRPS": 0.313,
+        "nMASE": 0.9851
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3118,
+        "nMASE": 0.9769
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6538,
+        "nMASE": 0.9843
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_14",
+      "term": "short",
+      "H": 30,
+      "items": 55,
+      "telos_small": {
+        "nCRPS": 0.4087,
+        "nMASE": 0.9362
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4035,
+        "nMASE": 0.9103
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6812,
+        "nMASE": 1.0087
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_15",
+      "term": "long",
+      "H": 180,
+      "items": 300,
+      "telos_small": {
+        "nCRPS": 0.176,
+        "nMASE": 0.841
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.174,
+        "nMASE": 0.7833
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4335,
+        "nMASE": 0.9553
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_15",
+      "term": "medium",
+      "H": 60,
+      "items": 300,
+      "telos_small": {
+        "nCRPS": 0.2548,
+        "nMASE": 1.0724
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2431,
+        "nMASE": 0.7703
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.4882,
+        "nMASE": 0.9309
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_15",
+      "term": "short",
+      "H": 30,
+      "items": 300,
+      "telos_small": {
+        "nCRPS": 0.3024,
+        "nMASE": 0.8212
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2996,
+        "nMASE": 0.7952
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5172,
+        "nMASE": 0.9246
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_16",
+      "term": "long",
+      "H": 180,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.3448,
+        "nMASE": 1.0914
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3489,
+        "nMASE": 1.0964
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5285,
+        "nMASE": 1.0157
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_16",
+      "term": "medium",
+      "H": 60,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.4802,
+        "nMASE": 1.1185
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.488,
+        "nMASE": 1.1406
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.608,
+        "nMASE": 1.0205
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_16",
+      "term": "short",
+      "H": 30,
+      "items": 10,
+      "telos_small": {
+        "nCRPS": 0.6118,
+        "nMASE": 1.1556
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.6265,
+        "nMASE": 1.1866
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6738,
+        "nMASE": 1.0917
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_17",
+      "term": "long",
+      "H": 180,
+      "items": 1020,
+      "telos_small": {
+        "nCRPS": 0.2281,
+        "nMASE": 0.9072
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2264,
+        "nMASE": 0.7981
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5017,
+        "nMASE": 0.8586
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_17",
+      "term": "medium",
+      "H": 60,
+      "items": 1020,
+      "telos_small": {
+        "nCRPS": 0.341,
+        "nMASE": 0.8718
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3359,
+        "nMASE": 0.7768
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6081,
+        "nMASE": 0.842
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_17",
+      "term": "short",
+      "H": 30,
+      "items": 1020,
+      "telos_small": {
+        "nCRPS": 0.4352,
+        "nMASE": 0.7983
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4239,
+        "nMASE": 0.7736
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6847,
+        "nMASE": 0.8347
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_18",
+      "term": "long",
+      "H": 180,
+      "items": 5,
+      "telos_small": {
+        "nCRPS": 0.3833,
+        "nMASE": 0.9783
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3824,
+        "nMASE": 0.9778
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7338,
+        "nMASE": 1.1111
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_18",
+      "term": "medium",
+      "H": 60,
+      "items": 5,
+      "telos_small": {
+        "nCRPS": 0.7491,
+        "nMASE": 1.135
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.55,
+        "nMASE": 0.9609
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7526,
+        "nMASE": 0.9451
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_18",
+      "term": "short",
+      "H": 30,
+      "items": 5,
+      "telos_small": {
+        "nCRPS": 0.5537,
+        "nMASE": 0.7949
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.551,
+        "nMASE": 0.794
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7254,
+        "nMASE": 0.839
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_19",
+      "term": "long",
+      "H": 180,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.3091,
+        "nMASE": 1.0488
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3093,
+        "nMASE": 0.7983
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6314,
+        "nMASE": 1.1146
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_19",
+      "term": "medium",
+      "H": 60,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.4712,
+        "nMASE": 1.1579
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4713,
+        "nMASE": 0.856
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7094,
+        "nMASE": 1.1134
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_19",
+      "term": "short",
+      "H": 30,
+      "items": 75,
+      "telos_small": {
+        "nCRPS": 0.5659,
+        "nMASE": 1.1666
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5656,
+        "nMASE": 0.8625
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7496,
+        "nMASE": 1.0855
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_20",
+      "term": "long",
+      "H": 180,
+      "items": 20,
+      "telos_small": {
+        "nCRPS": 0.2148,
+        "nMASE": 0.8507
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2148,
+        "nMASE": 0.8509
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6101,
+        "nMASE": 0.922
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_20",
+      "term": "medium",
+      "H": 60,
+      "items": 20,
+      "telos_small": {
+        "nCRPS": 0.3182,
+        "nMASE": 0.8547
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3136,
+        "nMASE": 0.8493
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6818,
+        "nMASE": 0.9231
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_20",
+      "term": "short",
+      "H": 30,
+      "items": 20,
+      "telos_small": {
+        "nCRPS": 0.3831,
+        "nMASE": 0.8272
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3818,
+        "nMASE": 0.826
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7244,
+        "nMASE": 0.9195
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_21",
+      "term": "long",
+      "H": 180,
+      "items": 240,
+      "telos_small": {
+        "nCRPS": 0.2462,
+        "nMASE": 0.8673
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.246,
+        "nMASE": 0.8549
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6614,
+        "nMASE": 0.9511
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_21",
+      "term": "medium",
+      "H": 60,
+      "items": 240,
+      "telos_small": {
+        "nCRPS": 0.3804,
+        "nMASE": 0.875
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3804,
+        "nMASE": 0.8752
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7294,
+        "nMASE": 0.9615
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_21",
+      "term": "short",
+      "H": 30,
+      "items": 240,
+      "telos_small": {
+        "nCRPS": 0.498,
+        "nMASE": 0.8733
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4976,
+        "nMASE": 0.8747
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7709,
+        "nMASE": 0.9727
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_22",
+      "term": "long",
+      "H": 180,
+      "items": 195,
+      "telos_small": {
+        "nCRPS": 0.3206,
+        "nMASE": 1.1734
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3088,
+        "nMASE": 0.8212
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.5181,
+        "nMASE": 0.9165
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_22",
+      "term": "medium",
+      "H": 60,
+      "items": 195,
+      "telos_small": {
+        "nCRPS": 0.4757,
+        "nMASE": 0.8722
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4734,
+        "nMASE": 0.8035
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6936,
+        "nMASE": 0.9035
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_22",
+      "term": "short",
+      "H": 30,
+      "items": 195,
+      "telos_small": {
+        "nCRPS": 0.5726,
+        "nMASE": 0.8218
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.5701,
+        "nMASE": 0.7898
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8024,
+        "nMASE": 0.8985
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_23",
+      "term": "long",
+      "H": 180,
+      "items": 865,
+      "telos_small": {
+        "nCRPS": 0.3037,
+        "nMASE": 0.9641
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2795,
+        "nMASE": 0.7822
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6584,
+        "nMASE": 0.8261
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_23",
+      "term": "medium",
+      "H": 60,
+      "items": 865,
+      "telos_small": {
+        "nCRPS": 0.3959,
+        "nMASE": 0.823
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3824,
+        "nMASE": 0.7557
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7111,
+        "nMASE": 0.8065
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_23",
+      "term": "short",
+      "H": 30,
+      "items": 865,
+      "telos_small": {
+        "nCRPS": 0.4754,
+        "nMASE": 0.7829
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4706,
+        "nMASE": 0.7516
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7408,
+        "nMASE": 0.7982
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_24",
+      "term": "long",
+      "H": 180,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.1805,
+        "nMASE": 1.4625
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.1626,
+        "nMASE": 0.7759
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.571,
+        "nMASE": 0.8651
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_24",
+      "term": "medium",
+      "H": 60,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.2729,
+        "nMASE": 1.0356
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2575,
+        "nMASE": 0.7599
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.6568,
+        "nMASE": 0.8796
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_24",
+      "term": "short",
+      "H": 30,
+      "items": 30,
+      "telos_small": {
+        "nCRPS": 0.3416,
+        "nMASE": 0.7422
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3412,
+        "nMASE": 0.7421
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7082,
+        "nMASE": 0.863
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_25",
+      "term": "long",
+      "H": 180,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.2767,
+        "nMASE": 0.8978
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.2763,
+        "nMASE": 0.7634
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7048,
+        "nMASE": 0.9026
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_25",
+      "term": "medium",
+      "H": 60,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.3989,
+        "nMASE": 0.7795
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.3988,
+        "nMASE": 0.7129
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.7689,
+        "nMASE": 0.876
+      }
+    },
+    {
+      "source": "tablet_compression",
+      "dataset": "code_25",
+      "term": "short",
+      "H": 30,
+      "items": 140,
+      "telos_small": {
+        "nCRPS": 0.4922,
+        "nMASE": 0.7184
+      },
+      "telos_small_pure_jepa": {
+        "nCRPS": 0.4928,
+        "nMASE": 0.7013
+      },
+      "analytical_baseline": {
+        "nCRPS": 0.8093,
+        "nMASE": 0.8634
+      }
+    }
+  ],
+  "model_family": "Telos",
+  "model_name": "Telos-Small"
+}

results/telos_small_training_report.json

@@ -0,0 +1,41 @@
+{
+  "model_name": "Telos-Small",
+  "model_tier": "Small",
+  "parameters": 50100747,
+  "architecture": {
+    "d_model": 512,
+    "num_heads": 8,
+    "encoder_blocks": 12,
+    "block_schedule": [
+      "time",
+      "time",
+      "time",
+      "time_variate",
+      "time",
+      "time",
+      "time",
+      "time_variate",
+      "time",
+      "time",
+      "time",
+      "time_variate"
+    ],
+    "jepa_predictor_blocks": 2,
+    "unet_skip_connections": true,
+    "value_residual_attention": true,
+    "single_shot_horizon_steps": 1024,
+    "eval_context_length": 4096,
+    "max_context_length": 8192,
+    "quantiles": 21
+  },
+  "training": {
+    "optimizer": "NorMuon (Newton-Schulz-5) + AdamW",
+    "steps": 43591,
+    "forecast_origins_seen": 12570406,
+    "skipped_steps": 0,
+    "elapsed_seconds": 29057.7,
+    "val_pinball_online": 0.13599,
+    "val_pinball_polyak": 0.13388,
+    "selected": "polyak"
+  }
+}

results/time_bench_telos_small_50m.json

@@ -0,0 +1,2139 @@
+{
+  "benchmark": "ICML-2026 `TIME` Benchmark (50 Fresh Datasets, 98 Tasks \u2014 100% Standalone Zero-Shot)",
+  "tasks_evaluated": 98,
+  "checkpoint_sha256": {
+    "telos_small": "ac7b7d2ff259aabdc7d94afccafae9790d2f9d3bb86097d16235ca10f7f2f7a7"
+  },
+  "parameters": 50100747,
+  "elapsed_s": 1980.7,
+  "leaderboard": [
+    {
+      "model_name": "Timestral-209M",
+      "provider": "Mistral AI",
+      "weights": "209M weights",
+      "norm_crps": 0.5223,
+      "crps_skill_pct": 47.77,
+      "norm_mase": 0.6262,
+      "mase_skill_pct": 37.38
+    },
+    {
+      "model_name": "Timestral-123M",
+      "provider": "Mistral AI",
+      "weights": "123M weights",
+      "norm_crps": 0.5253,
+      "crps_skill_pct": 47.47,
+      "norm_mase": 0.6277,
+      "mase_skill_pct": 37.23
+    },
+    {
+      "model_name": "t0-gamma",
+      "provider": "The Forecasting Co.",
+      "weights": "256M weights",
+      "norm_crps": 0.534,
+      "crps_skill_pct": 46.6,
+      "norm_mase": 0.637,
+      "mase_skill_pct": 36.3
+    },
+    {
+      "model_name": "Timestral-28M",
+      "provider": "Mistral AI",
+      "weights": "28M weights",
+      "norm_crps": 0.5344,
+      "crps_skill_pct": 46.56,
+      "norm_mase": 0.6382,
+      "mase_skill_pct": 36.18
+    },
+    {
+      "model_name": "QiYao-M",
+      "provider": "QiYao",
+      "weights": "Foundation",
+      "norm_crps": 0.5362,
+      "crps_skill_pct": 46.38,
+      "norm_mase": 0.6383,
+      "mase_skill_pct": 36.17
+    },
+    {
+      "model_name": "TimesFM-3",
+      "provider": "Google",
+      "weights": "331M weights",
+      "norm_crps": 0.5363,
+      "crps_skill_pct": 46.37,
+      "norm_mase": 0.6398,
+      "mase_skill_pct": 36.02
+    },
+    {
+      "model_name": "Toto-2.0 (2.5B)",
+      "provider": "Datadog",
+      "weights": "2.5B weights",
+      "norm_crps": 0.5394,
+      "crps_skill_pct": 46.06,
+      "norm_mase": 0.6419,
+      "mase_skill_pct": 35.81
+    },
+    {
+      "model_name": "Timestral-5M",
+      "provider": "Mistral AI",
+      "weights": "5.2M weights",
+      "norm_crps": 0.5445,
+      "crps_skill_pct": 45.55,
+      "norm_mase": 0.6518,
+      "mase_skill_pct": 34.82
+    },
+    {
+      "model_name": "Tabby",
+      "provider": "Tabby",
+      "weights": "147M weights",
+      "norm_crps": 0.5531,
+      "crps_skill_pct": 44.69,
+      "norm_mase": 0.6684,
+      "mase_skill_pct": 33.16
+    },
+    {
+      "model_name": "Chronos-2",
+      "provider": "AWS",
+      "weights": "120M weights",
+      "norm_crps": 0.5563,
+      "crps_skill_pct": 44.37,
+      "norm_mase": 0.662,
+      "mase_skill_pct": 33.8
+    },
+    {
+      "model_name": "TimesFM-2.5",
+      "provider": "Google",
+      "weights": "200M weights",
+      "norm_crps": 0.5674,
+      "crps_skill_pct": 43.26,
+      "norm_mase": 0.6686,
+      "mase_skill_pct": 33.14
+    },
+    {
+      "model_name": "TiRex",
+      "provider": "NX-AI",
+      "weights": "35M weights",
+      "norm_crps": 0.5731,
+      "crps_skill_pct": 42.69,
+      "norm_mase": 0.6831,
+      "mase_skill_pct": 31.69
+    },
+    {
+      "model_name": "Moirai-2.0",
+      "provider": "Salesforce",
+      "weights": "311M weights",
+      "norm_crps": 0.5885,
+      "crps_skill_pct": 41.15,
+      "norm_mase": 0.703,
+      "mase_skill_pct": 29.7
+    },
+    {
+      "model_name": "Chronos-Bolt",
+      "provider": "AWS",
+      "weights": "205M weights",
+      "norm_crps": 0.62,
+      "crps_skill_pct": 38.0,
+      "norm_mase": 0.7328,
+      "mase_skill_pct": 26.72
+    },
+    {
+      "model_name": "Telos-Small (50.1M Wts)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights (Standalone JEPA)",
+      "norm_crps": 0.643,
+      "crps_skill_pct": 35.7,
+      "norm_mase": 0.7616,
+      "mase_skill_pct": 23.84
+    },
+    {
+      "model_name": "Sundial-base",
+      "provider": "THUML",
+      "weights": "128M weights",
+      "norm_crps": 0.6634,
+      "crps_skill_pct": 33.66,
+      "norm_mase": 0.7576,
+      "mase_skill_pct": 24.24
+    },
+    {
+      "model_name": "Seasonal Naive",
+      "provider": "Baseline",
+      "weights": "0 weights",
+      "norm_crps": 1.0,
+      "crps_skill_pct": 0.0,
+      "norm_mase": 1.0,
+      "mase_skill_pct": 0.0
+    }
+  ],
+  "pure_nn_leaderboard": [
+    {
+      "model_name": "Telos-Small (Pure JEPA)",
+      "provider": "MAGI Labs",
+      "weights": "50.1M weights",
+      "norm_crps": 0.6346,
+      "crps_skill_pct": 36.54,
+      "norm_mase": 0.7672,
+      "mase_skill_pct": 23.28
+    }
+  ],
+  "per_task": [
+    {
+      "task": "Water_Quality_Darwin/15T/short",
+      "H": 16,
+      "V": 6,
+      "pairs": 3780,
+      "sn_mase": 1.1069051360803819,
+      "sn_crps": 0.10590079045606071,
+      "telos_small": {
+        "norm_crps": 0.5797,
+        "norm_mase": 0.8666
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5757,
+        "norm_mase": 0.916
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7859,
+        "norm_mase": 0.9977
+      }
+    },
+    {
+      "task": "Water_Quality_Darwin/15T/medium",
+      "H": 96,
+      "V": 6,
+      "pairs": 630,
+      "sn_mase": 1.10762925383268,
+      "sn_crps": 0.0994374783568173,
+      "telos_small": {
+        "norm_crps": 0.7502,
+        "norm_mase": 1.2578
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7455,
+        "norm_mase": 1.427
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8846,
+        "norm_mase": 1.1225
+      }
+    },
+    {
+      "task": "Water_Quality_Darwin/15T/long",
+      "H": 288,
+      "V": 6,
+      "pairs": 210,
+      "sn_mase": 1.6291144133574038,
+      "sn_crps": 0.13642779037572506,
+      "telos_small": {
+        "norm_crps": 0.7673,
+        "norm_mase": 1.2895
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7318,
+        "norm_mase": 1.3825
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8374,
+        "norm_mase": 1.1066
+      }
+    },
+    {
+      "task": "current_velocity/5T/short",
+      "H": 36,
+      "V": 6,
+      "pairs": 720,
+      "sn_mase": 0.9322325067949334,
+      "sn_crps": 0.45899238052691593,
+      "telos_small": {
+        "norm_crps": 0.4993,
+        "norm_mase": 0.4602
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.4994,
+        "norm_mase": 0.46
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7252,
+        "norm_mase": 0.8875
+      }
+    },
+    {
+      "task": "current_velocity/5T/medium",
+      "H": 288,
+      "V": 6,
+      "pairs": 90,
+      "sn_mase": 0.9184532526344081,
+      "sn_crps": 0.39367661529413084,
+      "telos_small": {
+        "norm_crps": 0.7024,
+        "norm_mase": 0.7935
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6975,
+        "norm_mase": 0.7637
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8025,
+        "norm_mase": 1.0096
+      }
+    },
+    {
+      "task": "current_velocity/5T/long",
+      "H": 864,
+      "V": 6,
+      "pairs": 30,
+      "sn_mase": 1.1509643237096372,
+      "sn_crps": 0.4297710248993255,
+      "telos_small": {
+        "norm_crps": 0.7372,
+        "norm_mase": 0.8561
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7056,
+        "norm_mase": 0.7989
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8083,
+        "norm_mase": 0.9667
+      }
+    },
+    {
+      "task": "current_velocity/10T/short",
+      "H": 18,
+      "V": 6,
+      "pairs": 7200,
+      "sn_mase": 1.054976530934313,
+      "sn_crps": 0.44623873889821497,
+      "telos_small": {
+        "norm_crps": 0.6347,
+        "norm_mase": 0.6582
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6357,
+        "norm_mase": 0.6577
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9075,
+        "norm_mase": 1.0485
+      }
+    },
+    {
+      "task": "current_velocity/10T/medium",
+      "H": 144,
+      "V": 6,
+      "pairs": 900,
+      "sn_mase": 1.0558433884087146,
+      "sn_crps": 0.3821147924837241,
+      "telos_small": {
+        "norm_crps": 0.793,
+        "norm_mase": 0.9505
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7908,
+        "norm_mase": 0.9577
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9839,
+        "norm_mase": 1.1448
+      }
+    },
+    {
+      "task": "current_velocity/10T/long",
+      "H": 432,
+      "V": 6,
+      "pairs": 300,
+      "sn_mase": 1.4548317170000027,
+      "sn_crps": 0.45246055181929196,
+      "telos_small": {
+        "norm_crps": 0.7568,
+        "norm_mase": 0.9114
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7398,
+        "norm_mase": 0.909
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9552,
+        "norm_mase": 1.0558
+      }
+    },
+    {
+      "task": "current_velocity/15T/short",
+      "H": 12,
+      "V": 6,
+      "pairs": 3600,
+      "sn_mase": 1.0051595530806665,
+      "sn_crps": 0.5436424048213736,
+      "telos_small": {
+        "norm_crps": 0.5634,
+        "norm_mase": 0.549
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5635,
+        "norm_mase": 0.549
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7444,
+        "norm_mase": 0.8241
+      }
+    },
+    {
+      "task": "current_velocity/15T/medium",
+      "H": 96,
+      "V": 6,
+      "pairs": 450,
+      "sn_mase": 1.0047960161367429,
+      "sn_crps": 0.4812549512047332,
+      "telos_small": {
+        "norm_crps": 0.7178,
+        "norm_mase": 0.813
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7162,
+        "norm_mase": 0.811
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8057,
+        "norm_mase": 0.9164
+      }
+    },
+    {
+      "task": "current_velocity/15T/long",
+      "H": 288,
+      "V": 6,
+      "pairs": 150,
+      "sn_mase": 1.138543533191249,
+      "sn_crps": 0.49196664824325503,
+      "telos_small": {
+        "norm_crps": 0.7538,
+        "norm_mase": 0.872
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7264,
+        "norm_mase": 0.8398
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8162,
+        "norm_mase": 0.9147
+      }
+    },
+    {
+      "task": "current_velocity/20T/short",
+      "H": 9,
+      "V": 6,
+      "pairs": 19440,
+      "sn_mase": 1.0270921800260557,
+      "sn_crps": 0.5368854185770512,
+      "telos_small": {
+        "norm_crps": 0.5297,
+        "norm_mase": 0.5737
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5342,
+        "norm_mase": 0.5774
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7955,
+        "norm_mase": 0.9031
+      }
+    },
+    {
+      "task": "current_velocity/20T/medium",
+      "H": 72,
+      "V": 6,
+      "pairs": 2430,
+      "sn_mase": 1.0316722379538157,
+      "sn_crps": 0.42822881608829516,
+      "telos_small": {
+        "norm_crps": 0.7696,
+        "norm_mase": 0.9088
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7853,
+        "norm_mase": 0.9413
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8753,
+        "norm_mase": 1.0018
+      }
+    },
+    {
+      "task": "current_velocity/20T/long",
+      "H": 216,
+      "V": 6,
+      "pairs": 810,
+      "sn_mase": 1.3419727755627,
+      "sn_crps": 0.4917035608531493,
+      "telos_small": {
+        "norm_crps": 0.7711,
+        "norm_mase": 0.9032
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7702,
+        "norm_mase": 0.9323
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8539,
+        "norm_mase": 0.966
+      }
+    },
+    {
+      "task": "current_velocity/H/short",
+      "H": 24,
+      "V": 6,
+      "pairs": 3528,
+      "sn_mase": 1.044176545302804,
+      "sn_crps": 0.37968425247374554,
+      "telos_small": {
+        "norm_crps": 0.817,
+        "norm_mase": 0.9806
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8118,
+        "norm_mase": 0.9615
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9697,
+        "norm_mase": 1.2064
+      }
+    },
+    {
+      "task": "current_velocity/H/medium",
+      "H": 168,
+      "V": 6,
+      "pairs": 504,
+      "sn_mase": 1.8636420656550583,
+      "sn_crps": 0.5224107098922609,
+      "telos_small": {
+        "norm_crps": 0.7198,
+        "norm_mase": 0.8495
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7123,
+        "norm_mase": 0.8509
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8126,
+        "norm_mase": 0.888
+      }
+    },
+    {
+      "task": "current_velocity/H/long",
+      "H": 336,
+      "V": 6,
+      "pairs": 252,
+      "sn_mase": 2.0181256953738695,
+      "sn_crps": 0.575336880615527,
+      "telos_small": {
+        "norm_crps": 0.6851,
+        "norm_mase": 0.8408
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6434,
+        "norm_mase": 0.8559
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7604,
+        "norm_mase": 0.8496
+      }
+    },
+    {
+      "task": "CPHL/15T/short",
+      "H": 12,
+      "V": 1,
+      "pairs": 240,
+      "sn_mase": 1.3920026617614616,
+      "sn_crps": 0.35606518946443716,
+      "telos_small": {
+        "norm_crps": 0.455,
+        "norm_mase": 0.501
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.4446,
+        "norm_mase": 0.4929
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7301,
+        "norm_mase": 0.7235
+      }
+    },
+    {
+      "task": "CPHL/15T/medium",
+      "H": 96,
+      "V": 1,
+      "pairs": 30,
+      "sn_mase": 1.3724792957278622,
+      "sn_crps": 0.30964362718345645,
+      "telos_small": {
+        "norm_crps": 0.7279,
+        "norm_mase": 0.7482
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7161,
+        "norm_mase": 0.7433
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8544,
+        "norm_mase": 0.8471
+      }
+    },
+    {
+      "task": "CPHL/15T/long",
+      "H": 288,
+      "V": 1,
+      "pairs": 10,
+      "sn_mase": 1.5632520412249478,
+      "sn_crps": 0.34880047791258295,
+      "telos_small": {
+        "norm_crps": 0.7128,
+        "norm_mase": 0.7379
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.693,
+        "norm_mase": 0.7155
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.79,
+        "norm_mase": 0.8071
+      }
+    },
+    {
+      "task": "CPHL/30T/short",
+      "H": 12,
+      "V": 1,
+      "pairs": 240,
+      "sn_mase": 1.7414882425655378,
+      "sn_crps": 0.3596878101498882,
+      "telos_small": {
+        "norm_crps": 0.6897,
+        "norm_mase": 0.7869
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.692,
+        "norm_mase": 0.8201
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7921,
+        "norm_mase": 0.8362
+      }
+    },
+    {
+      "task": "CPHL/30T/medium",
+      "H": 48,
+      "V": 1,
+      "pairs": 60,
+      "sn_mase": 1.6906705384673968,
+      "sn_crps": 0.3424080970435143,
+      "telos_small": {
+        "norm_crps": 0.7966,
+        "norm_mase": 0.8766
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.802,
+        "norm_mase": 0.9112
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8475,
+        "norm_mase": 0.8937
+      }
+    },
+    {
+      "task": "CPHL/30T/long",
+      "H": 144,
+      "V": 1,
+      "pairs": 20,
+      "sn_mase": 2.0687558533294763,
+      "sn_crps": 0.37329419606159353,
+      "telos_small": {
+        "norm_crps": 0.8616,
+        "norm_mase": 0.8608
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8552,
+        "norm_mase": 0.8688
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8788,
+        "norm_mase": 0.8655
+      }
+    },
+    {
+      "task": "CPHL/H/short",
+      "H": 24,
+      "V": 1,
+      "pairs": 112,
+      "sn_mase": 1.5151437641284222,
+      "sn_crps": 0.29740283600290723,
+      "telos_small": {
+        "norm_crps": 0.7641,
+        "norm_mase": 0.8295
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7628,
+        "norm_mase": 0.8357
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8522,
+        "norm_mase": 0.874
+      }
+    },
+    {
+      "task": "CPHL/H/medium",
+      "H": 168,
+      "V": 1,
+      "pairs": 16,
+      "sn_mase": 2.190084389033895,
+      "sn_crps": 0.39775350000876925,
+      "telos_small": {
+        "norm_crps": 0.8479,
+        "norm_mase": 0.89
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8054,
+        "norm_mase": 0.8663
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8653,
+        "norm_mase": 0.8757
+      }
+    },
+    {
+      "task": "CPHL/H/long",
+      "H": 336,
+      "V": 1,
+      "pairs": 8,
+      "sn_mase": 2.465174784023203,
+      "sn_crps": 0.4411259716788424,
+      "telos_small": {
+        "norm_crps": 0.8032,
+        "norm_mase": 0.8579
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7861,
+        "norm_mase": 0.8456
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8391,
+        "norm_mase": 0.8684
+      }
+    },
+    {
+      "task": "Coastal_T_S/5T/short",
+      "H": 36,
+      "V": 3,
+      "pairs": 6480,
+      "sn_mase": 1.2757414593343146,
+      "sn_crps": 0.08283715123506455,
+      "telos_small": {
+        "norm_crps": 0.5137,
+        "norm_mase": 0.6945
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5137,
+        "norm_mase": 0.6944
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7024,
+        "norm_mase": 0.8872
+      }
+    },
+    {
+      "task": "Coastal_T_S/5T/medium",
+      "H": 288,
+      "V": 3,
+      "pairs": 810,
+      "sn_mase": 1.2871949437162231,
+      "sn_crps": 0.08167608501346135,
+      "telos_small": {
+        "norm_crps": 0.5842,
+        "norm_mase": 0.9091
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5791,
+        "norm_mase": 0.9145
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7382,
+        "norm_mase": 0.9463
+      }
+    },
+    {
+      "task": "Coastal_T_S/5T/long",
+      "H": 864,
+      "V": 3,
+      "pairs": 270,
+      "sn_mase": 1.5473929157915876,
+      "sn_crps": 0.09199241843856108,
+      "telos_small": {
+        "norm_crps": 0.6081,
+        "norm_mase": 0.9707
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5855,
+        "norm_mase": 0.997
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7187,
+        "norm_mase": 0.9693
+      }
+    },
+    {
+      "task": "Coastal_T_S/15T/short",
+      "H": 12,
+      "V": 3,
+      "pairs": 1800,
+      "sn_mase": 1.486792847075776,
+      "sn_crps": 0.10880969790101082,
+      "telos_small": {
+        "norm_crps": 0.6712,
+        "norm_mase": 0.7706
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.669,
+        "norm_mase": 0.7705
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9409,
+        "norm_mase": 0.8534
+      }
+    },
+    {
+      "task": "Coastal_T_S/15T/medium",
+      "H": 96,
+      "V": 3,
+      "pairs": 225,
+      "sn_mase": 1.4931611681532413,
+      "sn_crps": 0.10701066296993984,
+      "telos_small": {
+        "norm_crps": 0.6862,
+        "norm_mase": 0.8532
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6743,
+        "norm_mase": 0.8547
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9561,
+        "norm_mase": 0.8843
+      }
+    },
+    {
+      "task": "Coastal_T_S/15T/long",
+      "H": 288,
+      "V": 3,
+      "pairs": 75,
+      "sn_mase": 1.6240005522592336,
+      "sn_crps": 0.110788018959756,
+      "telos_small": {
+        "norm_crps": 0.6661,
+        "norm_mase": 0.8634
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6568,
+        "norm_mase": 0.8665
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9187,
+        "norm_mase": 0.8885
+      }
+    },
+    {
+      "task": "Coastal_T_S/20T/short",
+      "H": 9,
+      "V": 3,
+      "pairs": 360,
+      "sn_mase": 1.270873201451952,
+      "sn_crps": 0.005451858318258226,
+      "telos_small": {
+        "norm_crps": 0.6104,
+        "norm_mase": 0.5796
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6147,
+        "norm_mase": 0.5848
+      },
+      "analytical_baseline": {
+        "norm_crps": 1.1998,
+        "norm_mase": 1.1031
+      }
+    },
+    {
+      "task": "Coastal_T_S/20T/medium",
+      "H": 72,
+      "V": 3,
+      "pairs": 45,
+      "sn_mase": 1.2910994490270333,
+      "sn_crps": 0.005468844758050755,
+      "telos_small": {
+        "norm_crps": 1.1501,
+        "norm_mase": 1.0739
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 1.2083,
+        "norm_mase": 1.1037
+      },
+      "analytical_baseline": {
+        "norm_crps": 1.3354,
+        "norm_mase": 1.1991
+      }
+    },
+    {
+      "task": "Coastal_T_S/20T/long",
+      "H": 216,
+      "V": 3,
+      "pairs": 15,
+      "sn_mase": 1.8180515700802744,
+      "sn_crps": 0.007914651740110096,
+      "telos_small": {
+        "norm_crps": 1.0869,
+        "norm_mase": 1.0519
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 1.1028,
+        "norm_mase": 1.0544
+      },
+      "analytical_baseline": {
+        "norm_crps": 1.2249,
+        "norm_mase": 1.1385
+      }
+    },
+    {
+      "task": "Coastal_T_S/H/short",
+      "H": 24,
+      "V": 3,
+      "pairs": 2016,
+      "sn_mase": 1.2444395848743397,
+      "sn_crps": 0.03564928423529185,
+      "telos_small": {
+        "norm_crps": 0.7669,
+        "norm_mase": 0.9901
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7871,
+        "norm_mase": 0.9706
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7972,
+        "norm_mase": 1.1695
+      }
+    },
+    {
+      "task": "Coastal_T_S/H/medium",
+      "H": 168,
+      "V": 3,
+      "pairs": 288,
+      "sn_mase": 2.121073907527163,
+      "sn_crps": 0.05280124069402554,
+      "telos_small": {
+        "norm_crps": 0.7412,
+        "norm_mase": 0.948
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7661,
+        "norm_mase": 0.9823
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7446,
+        "norm_mase": 0.935
+      }
+    },
+    {
+      "task": "Coastal_T_S/H/long",
+      "H": 336,
+      "V": 3,
+      "pairs": 144,
+      "sn_mase": 2.7461676211938584,
+      "sn_crps": 0.06265422569084722,
+      "telos_small": {
+        "norm_crps": 0.6389,
+        "norm_mase": 0.9323
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6799,
+        "norm_mase": 0.9943
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6617,
+        "norm_mase": 0.9635
+      }
+    },
+    {
+      "task": "SG_Weather/D/short",
+      "H": 3,
+      "V": 4,
+      "pairs": 2928,
+      "sn_mase": 1.093336152628154,
+      "sn_crps": 0.2978323671428984,
+      "telos_small": {
+        "norm_crps": 0.6825,
+        "norm_mase": 0.8063
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6765,
+        "norm_mase": 0.8073
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7622,
+        "norm_mase": 0.8277
+      }
+    },
+    {
+      "task": "SG_Weather/D/medium",
+      "H": 7,
+      "V": 4,
+      "pairs": 1248,
+      "sn_mase": 1.2176871928440223,
+      "sn_crps": 0.319348745939327,
+      "telos_small": {
+        "norm_crps": 0.5864,
+        "norm_mase": 0.7637
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5829,
+        "norm_mase": 0.7636
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6626,
+        "norm_mase": 0.7962
+      }
+    },
+    {
+      "task": "SG_Weather/D/long",
+      "H": 14,
+      "V": 4,
+      "pairs": 624,
+      "sn_mase": 1.271157955343153,
+      "sn_crps": 0.3508238380771969,
+      "telos_small": {
+        "norm_crps": 0.5238,
+        "norm_mase": 0.7433
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5197,
+        "norm_mase": 0.7467
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6283,
+        "norm_mase": 0.7617
+      }
+    },
+    {
+      "task": "SG_PM25/H/short",
+      "H": 24,
+      "V": 5,
+      "pairs": 460,
+      "sn_mase": 1.1786627570510653,
+      "sn_crps": 0.4301059762756827,
+      "telos_small": {
+        "norm_crps": 0.7075,
+        "norm_mase": 0.7414
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7044,
+        "norm_mase": 0.7388
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7818,
+        "norm_mase": 0.7931
+      }
+    },
+    {
+      "task": "SG_PM25/H/medium",
+      "H": 72,
+      "V": 5,
+      "pairs": 150,
+      "sn_mase": 1.210027230240395,
+      "sn_crps": 0.4387095673373417,
+      "telos_small": {
+        "norm_crps": 0.7133,
+        "norm_mase": 0.7613
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7126,
+        "norm_mase": 0.7624
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7793,
+        "norm_mase": 0.8015
+      }
+    },
+    {
+      "task": "SG_PM25/H/long",
+      "H": 168,
+      "V": 5,
+      "pairs": 65,
+      "sn_mase": 1.3545130367993115,
+      "sn_crps": 0.5191195690792656,
+      "telos_small": {
+        "norm_crps": 0.6214,
+        "norm_mase": 0.7071
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6147,
+        "norm_mase": 0.7005
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.69,
+        "norm_mase": 0.7487
+      }
+    },
+    {
+      "task": "NE_China_Wind/H/short",
+      "H": 24,
+      "V": 4,
+      "pairs": 120,
+      "sn_mase": 0.8637674135993827,
+      "sn_crps": 0.620670741528828,
+      "telos_small": {
+        "norm_crps": 0.5793,
+        "norm_mase": 0.6416
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5749,
+        "norm_mase": 0.639
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8112,
+        "norm_mase": 0.9276
+      }
+    },
+    {
+      "task": "NE_China_Wind/H/medium",
+      "H": 72,
+      "V": 4,
+      "pairs": 40,
+      "sn_mase": 1.1205075075975954,
+      "sn_crps": 0.6438977838477045,
+      "telos_small": {
+        "norm_crps": 0.7165,
+        "norm_mase": 0.8778
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7029,
+        "norm_mase": 0.8681
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8464,
+        "norm_mase": 0.9442
+      }
+    },
+    {
+      "task": "NE_China_Wind/H/long",
+      "H": 168,
+      "V": 4,
+      "pairs": 16,
+      "sn_mase": 1.4626812861088365,
+      "sn_crps": 0.7264193737897133,
+      "telos_small": {
+        "norm_crps": 0.6532,
+        "norm_mase": 0.852
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6435,
+        "norm_mase": 0.8206
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7747,
+        "norm_mase": 0.8907
+      }
+    },
+    {
+      "task": "Australia_Solar/H/short",
+      "H": 24,
+      "V": 3,
+      "pairs": 315,
+      "sn_mase": 0.6966077060819641,
+      "sn_crps": 0.3164510454764971,
+      "telos_small": {
+        "norm_crps": 0.9257,
+        "norm_mase": 0.9035
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 1.0073,
+        "norm_mase": 0.9926
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9471,
+        "norm_mase": 0.9367
+      }
+    },
+    {
+      "task": "Australia_Solar/H/medium",
+      "H": 72,
+      "V": 3,
+      "pairs": 105,
+      "sn_mase": 0.7543073106302655,
+      "sn_crps": 0.3078652232179442,
+      "telos_small": {
+        "norm_crps": 0.7927,
+        "norm_mase": 0.827
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8256,
+        "norm_mase": 0.865
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8581,
+        "norm_mase": 0.891
+      }
+    },
+    {
+      "task": "Australia_Solar/H/long",
+      "H": 168,
+      "V": 3,
+      "pairs": 45,
+      "sn_mase": 0.6799651516183869,
+      "sn_crps": 0.3093677736752417,
+      "telos_small": {
+        "norm_crps": 0.7165,
+        "norm_mase": 0.8791
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7333,
+        "norm_mase": 0.9015
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8325,
+        "norm_mase": 0.9262
+      }
+    },
+    {
+      "task": "epf_electricity_price/H/short",
+      "H": 24,
+      "V": 1,
+      "pairs": 525,
+      "sn_mase": 1.4417132029177866,
+      "sn_crps": 0.19760442070876835,
+      "telos_small": {
+        "norm_crps": 0.6986,
+        "norm_mase": 0.7243
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6968,
+        "norm_mase": 0.7219
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8082,
+        "norm_mase": 0.8248
+      }
+    },
+    {
+      "task": "epf_electricity_price/H/medium",
+      "H": 72,
+      "V": 1,
+      "pairs": 175,
+      "sn_mase": 1.8714812741253186,
+      "sn_crps": 0.2429715270865831,
+      "telos_small": {
+        "norm_crps": 0.6511,
+        "norm_mase": 0.6925
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6472,
+        "norm_mase": 0.6885
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7682,
+        "norm_mase": 0.8028
+      }
+    },
+    {
+      "task": "epf_electricity_price/H/long",
+      "H": 168,
+      "V": 1,
+      "pairs": 75,
+      "sn_mase": 2.0093576546971135,
+      "sn_crps": 0.2617513494932196,
+      "telos_small": {
+        "norm_crps": 0.6471,
+        "norm_mase": 0.7304
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6324,
+        "norm_mase": 0.7187
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.751,
+        "norm_mase": 0.8089
+      }
+    },
+    {
+      "task": "OpenElectricity_NEM/5T/short",
+      "H": 24,
+      "V": 10,
+      "pairs": 1680,
+      "sn_mase": 1.0344248622662509,
+      "sn_crps": 0.4122996004505975,
+      "telos_small": {
+        "norm_crps": 0.3746,
+        "norm_mase": 0.4548
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.3662,
+        "norm_mase": 0.4508
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6395,
+        "norm_mase": 0.6416
+      }
+    },
+    {
+      "task": "OpenElectricity_NEM/5T/medium",
+      "H": 96,
+      "V": 10,
+      "pairs": 420,
+      "sn_mase": 1.0517636814146012,
+      "sn_crps": 0.34281670960204264,
+      "telos_small": {
+        "norm_crps": 0.614,
+        "norm_mase": 0.6728
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.597,
+        "norm_mase": 0.6747
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7143,
+        "norm_mase": 0.7442
+      }
+    },
+    {
+      "task": "OpenElectricity_NEM/5T/long",
+      "H": 288,
+      "V": 10,
+      "pairs": 140,
+      "sn_mase": 1.0447330270897144,
+      "sn_crps": 0.26569570571267215,
+      "telos_small": {
+        "norm_crps": 0.8026,
+        "norm_mase": 0.7873
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8068,
+        "norm_mase": 0.7961
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8276,
+        "norm_mase": 0.8362
+      }
+    },
+    {
+      "task": "EWELD_Load/15T/short",
+      "H": 24,
+      "V": 10,
+      "pairs": 560,
+      "sn_mase": 0.9696418375430754,
+      "sn_crps": 2.2708062592940985,
+      "telos_small": {
+        "norm_crps": 0.3367,
+        "norm_mase": 0.6855
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.2476,
+        "norm_mase": 0.7467
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.3355,
+        "norm_mase": 0.7123
+      }
+    },
+    {
+      "task": "EWELD_Load/15T/medium",
+      "H": 96,
+      "V": 10,
+      "pairs": 140,
+      "sn_mase": 0.946050857924023,
+      "sn_crps": 0.3261333688739633,
+      "telos_small": {
+        "norm_crps": 0.6309,
+        "norm_mase": 0.7748
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.674,
+        "norm_mase": 0.8605
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7163,
+        "norm_mase": 0.8001
+      }
+    },
+    {
+      "task": "EWELD_Load/15T/long",
+      "H": 672,
+      "V": 10,
+      "pairs": 20,
+      "sn_mase": 1.0387980045077339,
+      "sn_crps": 0.3752873326349839,
+      "telos_small": {
+        "norm_crps": 0.5551,
+        "norm_mase": 0.8324
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.584,
+        "norm_mase": 0.9105
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6187,
+        "norm_mase": 0.8035
+      }
+    },
+    {
+      "task": "SG_Carpark/15T/short",
+      "H": 16,
+      "V": 1,
+      "pairs": 14868,
+      "sn_mase": 0.9522074667547793,
+      "sn_crps": 0.0781783317710341,
+      "telos_small": {
+        "norm_crps": 0.5459,
+        "norm_mase": 0.5625
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5214,
+        "norm_mase": 0.5521
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7192,
+        "norm_mase": 0.7099
+      }
+    },
+    {
+      "task": "SG_Carpark/15T/medium",
+      "H": 96,
+      "V": 1,
+      "pairs": 2478,
+      "sn_mase": 0.9573854787444838,
+      "sn_crps": 0.07004291432227401,
+      "telos_small": {
+        "norm_crps": 0.7495,
+        "norm_mase": 0.7569
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.78,
+        "norm_mase": 0.8346
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9075,
+        "norm_mase": 0.9278
+      }
+    },
+    {
+      "task": "SG_Carpark/15T/long",
+      "H": 672,
+      "V": 1,
+      "pairs": 354,
+      "sn_mase": 1.3412298857390095,
+      "sn_crps": 0.10256839424830191,
+      "telos_small": {
+        "norm_crps": 0.6187,
+        "norm_mase": 0.6492
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6074,
+        "norm_mase": 0.6773
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7197,
+        "norm_mase": 0.7566
+      }
+    },
+    {
+      "task": "Finland_Traffic/15T/short",
+      "H": 16,
+      "V": 1,
+      "pairs": 186,
+      "sn_mase": 0.831073451998675,
+      "sn_crps": 0.5190304420163312,
+      "telos_small": {
+        "norm_crps": 0.5139,
+        "norm_mase": 0.6412
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5937,
+        "norm_mase": 0.6899
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6522,
+        "norm_mase": 0.7692
+      }
+    },
+    {
+      "task": "Finland_Traffic/15T/medium",
+      "H": 96,
+      "V": 1,
+      "pairs": 31,
+      "sn_mase": 0.7899321246773803,
+      "sn_crps": 0.2426313677259669,
+      "telos_small": {
+        "norm_crps": 0.6523,
+        "norm_mase": 0.7355
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7367,
+        "norm_mase": 0.8311
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8276,
+        "norm_mase": 0.9091
+      }
+    },
+    {
+      "task": "Finland_Traffic/15T/long",
+      "H": 672,
+      "V": 1,
+      "pairs": 4,
+      "sn_mase": 1.2219304045711639,
+      "sn_crps": 0.3688619414773533,
+      "telos_small": {
+        "norm_crps": 0.438,
+        "norm_mase": 0.4934
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.4749,
+        "norm_mase": 0.5474
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.5844,
+        "norm_mase": 0.6362
+      }
+    },
+    {
+      "task": "Port_Activity/D/short",
+      "H": 30,
+      "V": 2,
+      "pairs": 2376,
+      "sn_mase": 1.1471287042428393,
+      "sn_crps": 3.361978570994329,
+      "telos_small": {
+        "norm_crps": 0.1795,
+        "norm_mase": 0.5795
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.1792,
+        "norm_mase": 0.5788
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.2647,
+        "norm_mase": 0.6382
+      }
+    },
+    {
+      "task": "Port_Activity/W/short",
+      "H": 13,
+      "V": 2,
+      "pairs": 792,
+      "sn_mase": 1.0297572925672263,
+      "sn_crps": 0.4940322955479891,
+      "telos_small": {
+        "norm_crps": 0.5328,
+        "norm_mase": 0.7551
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5329,
+        "norm_mase": 0.7547
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7127,
+        "norm_mase": 0.7788
+      }
+    },
+    {
+      "task": "ECDC_COVID/D/short",
+      "H": 30,
+      "V": 1,
+      "pairs": 45,
+      "sn_mase": 3.5775833289982253,
+      "sn_crps": 0.6199984789319984,
+      "telos_small": {
+        "norm_crps": 0.8762,
+        "norm_mase": 1.0586
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.915,
+        "norm_mase": 1.039
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8852,
+        "norm_mase": 1.0093
+      }
+    },
+    {
+      "task": "ECDC_COVID/W/short",
+      "H": 13,
+      "V": 1,
+      "pairs": 64,
+      "sn_mase": 1.624622403319587,
+      "sn_crps": 0.7955525926169933,
+      "telos_small": {
+        "norm_crps": 0.9039,
+        "norm_mase": 1.0779
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.9052,
+        "norm_mase": 1.1126
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9424,
+        "norm_mase": 1.0273
+      }
+    },
+    {
+      "task": "Global_Influenza/W/short",
+      "H": 13,
+      "V": 4,
+      "pairs": 240,
+      "sn_mase": 4.326502407633893,
+      "sn_crps": 1.3337073250013445,
+      "telos_small": {
+        "norm_crps": 0.4884,
+        "norm_mase": 0.6916
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.3204,
+        "norm_mase": 0.6509
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6878,
+        "norm_mase": 0.9034
+      }
+    },
+    {
+      "task": "Crypto/D/short",
+      "H": 30,
+      "V": 4,
+      "pairs": 36,
+      "sn_mase": 6.299691296108334,
+      "sn_crps": 0.08319199341598717,
+      "telos_small": {
+        "norm_crps": 0.9163,
+        "norm_mase": 1.0301
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.9122,
+        "norm_mase": 1.0958
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9025,
+        "norm_mase": 0.9846
+      }
+    },
+    {
+      "task": "US_Term_Structure/B/short",
+      "H": 20,
+      "V": 40,
+      "pairs": 1400,
+      "sn_mase": 1.7412022422181404,
+      "sn_crps": 0.248434916420281,
+      "telos_small": {
+        "norm_crps": 0.8931,
+        "norm_mase": 0.9213
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8358,
+        "norm_mase": 0.9047
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9398,
+        "norm_mase": 0.9781
+      }
+    },
+    {
+      "task": "Oil_Price/B/short",
+      "H": 20,
+      "V": 12,
+      "pairs": 420,
+      "sn_mase": 1.5005212057649062,
+      "sn_crps": 0.053385782034238397,
+      "telos_small": {
+        "norm_crps": 0.8623,
+        "norm_mase": 0.9074
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.8607,
+        "norm_mase": 0.9028
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9463,
+        "norm_mase": 0.9805
+      }
+    },
+    {
+      "task": "Job_Claims/W/short",
+      "H": 13,
+      "V": 2,
+      "pairs": 8,
+      "sn_mase": 2.845079840499296,
+      "sn_crps": 0.019447086324874976,
+      "telos_small": {
+        "norm_crps": 0.9556,
+        "norm_mase": 1.0505
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 1.0757,
+        "norm_mase": 1.3753
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.9858,
+        "norm_mase": 0.942
+      }
+    },
+    {
+      "task": "Uncertainty_1M/M/short",
+      "H": 6,
+      "V": 3,
+      "pairs": 21,
+      "sn_mase": 1.057881699564403,
+      "sn_crps": 0.07935248774802403,
+      "telos_small": {
+        "norm_crps": 0.3571,
+        "norm_mase": 0.3321
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.3417,
+        "norm_mase": 0.3183
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.6141,
+        "norm_mase": 0.5698
+      }
+    },
+    {
+      "task": "Housing_Inventory/M/short",
+      "H": 12,
+      "V": 4,
+      "pairs": 12,
+      "sn_mase": 0.7186913127542733,
+      "sn_crps": 0.0665798563186104,
+      "telos_small": {
+        "norm_crps": 0.6062,
+        "norm_mase": 0.652
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6054,
+        "norm_mase": 0.6613
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8638,
+        "norm_mase": 0.8614
+      }
+    },
+    {
+      "task": "JOLTS/M/short",
+      "H": 12,
+      "V": 6,
+      "pairs": 30,
+      "sn_mase": 1.633600374062145,
+      "sn_crps": 0.10072316522205195,
+      "telos_small": {
+        "norm_crps": 0.8847,
+        "norm_mase": 0.8779
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7449,
+        "norm_mase": 0.7685
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7784,
+        "norm_mase": 0.7941
+      }
+    },
+    {
+      "task": "US_Labor/M/short",
+      "H": 12,
+      "V": 14,
+      "pairs": 70,
+      "sn_mase": 1.7849303193866493,
+      "sn_crps": 0.11137483884199327,
+      "telos_small": {
+        "norm_crps": 0.5952,
+        "norm_mase": 0.6249
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.4832,
+        "norm_mase": 0.5078
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7294,
+        "norm_mase": 0.7442
+      }
+    },
+    {
+      "task": "Vehicle_Supply/M/short",
+      "H": 12,
+      "V": 6,
+      "pairs": 30,
+      "sn_mase": 1.1103281271782204,
+      "sn_crps": 0.34748026529032994,
+      "telos_small": {
+        "norm_crps": 0.5923,
+        "norm_mase": 0.7318
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5407,
+        "norm_mase": 0.7006
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.884,
+        "norm_mase": 0.8821
+      }
+    },
+    {
+      "task": "Auto_Production_SF/M/short",
+      "H": 12,
+      "V": 1,
+      "pairs": 5,
+      "sn_mase": 0.977221482631875,
+      "sn_crps": 0.031800960373476926,
+      "telos_small": {
+        "norm_crps": 0.8685,
+        "norm_mase": 0.9429
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.9639,
+        "norm_mase": 0.9968
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8799,
+        "norm_mase": 0.8964
+      }
+    },
+    {
+      "task": "Commodity_Production/M/short",
+      "H": 12,
+      "V": 1,
+      "pairs": 160,
+      "sn_mase": 1.2460988308860295,
+      "sn_crps": 0.1777932016921801,
+      "telos_small": {
+        "norm_crps": 0.7058,
+        "norm_mase": 0.7745
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.701,
+        "norm_mase": 0.7704
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7952,
+        "norm_mase": 0.8661
+      }
+    },
+    {
+      "task": "Commodity_Import/M/short",
+      "H": 12,
+      "V": 1,
+      "pairs": 40,
+      "sn_mase": 1.866751766074178,
+      "sn_crps": 0.33271580939241,
+      "telos_small": {
+        "norm_crps": 0.6547,
+        "norm_mase": 0.8736
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6541,
+        "norm_mase": 0.8708
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7451,
+        "norm_mase": 0.9166
+      }
+    },
+    {
+      "task": "WUI_Global/Q/short",
+      "H": 4,
+      "V": 15,
+      "pairs": 75,
+      "sn_mase": 1.3394226477728888,
+      "sn_crps": 0.8246595045704057,
+      "telos_small": {
+        "norm_crps": 0.6634,
+        "norm_mase": 0.7364
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.627,
+        "norm_mase": 0.7119
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7634,
+        "norm_mase": 0.8303
+      }
+    },
+    {
+      "task": "Global_Price/Q/short",
+      "H": 4,
+      "V": 60,
+      "pairs": 300,
+      "sn_mase": 1.8238301725900739,
+      "sn_crps": 0.1794111687923069,
+      "telos_small": {
+        "norm_crps": 0.8342,
+        "norm_mase": 0.8701
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.804,
+        "norm_mase": 0.8111
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8975,
+        "norm_mase": 0.9141
+      }
+    },
+    {
+      "task": "Vehicle_Sales/M/short",
+      "H": 12,
+      "V": 10,
+      "pairs": 50,
+      "sn_mase": 1.4095663160270906,
+      "sn_crps": 0.11473660792759578,
+      "telos_small": {
+        "norm_crps": 0.6279,
+        "norm_mase": 0.6494
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6013,
+        "norm_mase": 0.608
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7852,
+        "norm_mase": 0.7432
+      }
+    },
+    {
+      "task": "Online_Retail_2_UCI/D/short",
+      "H": 30,
+      "V": 1,
+      "pairs": 6,
+      "sn_mase": 2.140780928301401,
+      "sn_crps": 0.999642567118059,
+      "telos_small": {
+        "norm_crps": 0.2751,
+        "norm_mase": 0.3346
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.2658,
+        "norm_mase": 0.3307
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.5166,
+        "norm_mase": 0.4906
+      }
+    },
+    {
+      "task": "Supply_Chain_Customer/D/short",
+      "H": 30,
+      "V": 36,
+      "pairs": 432,
+      "sn_mase": 1.6936774961938938,
+      "sn_crps": 1.0689042573613643,
+      "telos_small": {
+        "norm_crps": 0.2884,
+        "norm_mase": 0.3898
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.2877,
+        "norm_mase": 0.3848
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.5402,
+        "norm_mase": 0.5964
+      }
+    },
+    {
+      "task": "Supply_Chain_Location/D/short",
+      "H": 30,
+      "V": 51,
+      "pairs": 612,
+      "sn_mase": 1.5971196207187819,
+      "sn_crps": 1.0577564930502144,
+      "telos_small": {
+        "norm_crps": 0.2973,
+        "norm_mase": 0.4248
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.2985,
+        "norm_mase": 0.4229
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.5554,
+        "norm_mase": 0.6297
+      }
+    },
+    {
+      "task": "azure2019_D/5T/short",
+      "H": 288,
+      "V": 3,
+      "pairs": 8901,
+      "sn_mase": 1.09388082352075,
+      "sn_crps": 0.2926142028628922,
+      "telos_small": {
+        "norm_crps": 0.614,
+        "norm_mase": 0.7869
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5956,
+        "norm_mase": 0.7761
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.712,
+        "norm_mase": 0.8441
+      }
+    },
+    {
+      "task": "azure2019_I/5T/short",
+      "H": 288,
+      "V": 3,
+      "pairs": 4428,
+      "sn_mase": 1.0115609729058312,
+      "sn_crps": 0.22805883596642443,
+      "telos_small": {
+        "norm_crps": 0.7147,
+        "norm_mase": 0.8116
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6287,
+        "norm_mase": 0.8111
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7438,
+        "norm_mase": 0.8323
+      }
+    },
+    {
+      "task": "azure2019_U/5T/short",
+      "H": 48,
+      "V": 3,
+      "pairs": 1404,
+      "sn_mase": 1.1517131263146043,
+      "sn_crps": 0.570542980609912,
+      "telos_small": {
+        "norm_crps": 0.3425,
+        "norm_mase": 0.6831
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.3372,
+        "norm_mase": 0.6787
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.5269,
+        "norm_mase": 0.8219
+      }
+    },
+    {
+      "task": "Smart_Manufacturing/H/short",
+      "H": 24,
+      "V": 5,
+      "pairs": 2380,
+      "sn_mase": 1.0055067714974981,
+      "sn_crps": 0.23542101313355804,
+      "telos_small": {
+        "norm_crps": 0.7095,
+        "norm_mase": 0.7097
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.7087,
+        "norm_mase": 0.7089
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.77,
+        "norm_mase": 0.7712
+      }
+    },
+    {
+      "task": "Smart_Manufacturing/H/medium",
+      "H": 168,
+      "V": 5,
+      "pairs": 340,
+      "sn_mase": 1.044595285129953,
+      "sn_crps": 0.27220983582984515,
+      "telos_small": {
+        "norm_crps": 0.61,
+        "norm_mase": 0.6831
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6097,
+        "norm_mase": 0.6827
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7453,
+        "norm_mase": 0.7438
+      }
+    },
+    {
+      "task": "Smart_Manufacturing/H/long",
+      "H": 336,
+      "V": 5,
+      "pairs": 170,
+      "sn_mase": 1.0719038232576714,
+      "sn_crps": 0.31738381102122837,
+      "telos_small": {
+        "norm_crps": 0.5229,
+        "norm_mase": 0.6655
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.5227,
+        "norm_mase": 0.6652
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7139,
+        "norm_mase": 0.7235
+      }
+    },
+    {
+      "task": "MetroPT-3/5T/short",
+      "H": 48,
+      "V": 6,
+      "pairs": 216,
+      "sn_mase": 0.8450006323641208,
+      "sn_crps": 0.37834817117371744,
+      "telos_small": {
+        "norm_crps": 0.691,
+        "norm_mase": 0.7594
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6885,
+        "norm_mase": 0.7571
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8309,
+        "norm_mase": 0.8569
+      }
+    },
+    {
+      "task": "MetroPT-3/5T/medium",
+      "H": 288,
+      "V": 6,
+      "pairs": 36,
+      "sn_mase": 0.8333397275589067,
+      "sn_crps": 0.36930167750747755,
+      "telos_small": {
+        "norm_crps": 0.7036,
+        "norm_mase": 0.7827
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6906,
+        "norm_mase": 0.7652
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.8286,
+        "norm_mase": 0.8658
+      }
+    },
+    {
+      "task": "MetroPT-3/5T/long",
+      "H": 576,
+      "V": 6,
+      "pairs": 18,
+      "sn_mase": 0.8545103668350666,
+      "sn_crps": 0.38544667902998536,
+      "telos_small": {
+        "norm_crps": 0.6866,
+        "norm_mase": 0.7925
+      },
+      "telos_small_pure_jepa": {
+        "norm_crps": 0.6661,
+        "norm_mase": 0.7584
+      },
+      "analytical_baseline": {
+        "norm_crps": 0.7962,
+        "norm_mase": 0.8375
+      }
+    }
+  ],
+  "model_family": "Telos",
+  "model_name": "Telos-Small"
+}