DangerLabs/dm-jepa-1.5

DM-JEPA 1.5: System 1 Predictive Decision Architecture

Organization: Danger Labs
Version: v1.5.0-jepa2
License: Danger Labs Open-Weights License (Open Weights / Closed Training Pipelines)
Parameter Count: 94.24 Million
Primary Format: model.safetensors (Zero-Code-Execution SafeTensors)

---

1. Release Overview & Policy

This repository provides the official Open-Weights release of DM-JEPA 1.5 (powered by Danger Labs' next-generation JEPA 2.0 Engine).

Policy Notice:
In accordance with Danger Labs governance, this release is strictly Open Weights (not Open Source). Pretrained parameter tensors, architecture configurations, and tokenization dictionaries are provided freely for empirical verification, academic research, and edge deployment. Danger Labs' proprietary multi-tier distillation recipes, synthetic scenario generators, and training infrastructure remain closed intellectual property.

---

2. Key Architectural Innovations

DM-JEPA 1.5 is a non-autoregressive, energy-based Joint-Embedding Predictive Architecture (JEPA) designed for high-frequency System 1 decision-making under strict real-time constraints:

1. Sub-Quadratic Cartesian Block Butterfly Factorization: Replaces conventional overparameterized Feed-Forward Networks with factorized 32x32 block matrices W = B1 * B2, cutting non-linear projection parameters by 75% while accelerating tensor throughput.
2. Riemannian Manifold Compatibility Scorer: Unlike isotropic Euclidean or cosine distance heads that fail when ranking subtle or conflicting action candidates (e.g. creep vs yield), the JEPA 2.0 scorer projects predicted latent states (s_hat) and candidate options (s_k) through a learned Riemannian metric tensor M = B1 * B2, enabling curved boundary separation and hierarchical decision arbitration.
3. Adaptive Latent Energy Early-Exit: Dynamically monitors latent thought delta (||Δs|| < ε). Deterministic and straightforward actions resolve in a single step (< 2 ms), reserving multi-step iterative rollouts exclusively for complex, multi-agent edge hazards.
4. First-Class Fail-Safe Candidate (NULL_ACTION_ABSTAIN): Directly parameterized candidate embedding providing zero-cost open-set refusal and ISO 26262 ASIL-D emergency stops when scenario trajectories violate safety bounds.

---

3. Verified Benchmark Results

A. Official Decision Index (Edition 0.3)

Evaluated across all 37 official benchmarks spanning 5 core areas on the canonical Decision Index Edition 0.3 test suite:

Model	Parameters	Public Decision Index	Generalization (v)	Balanced Raw	Coverage	Forward Latency
DM-JEPA 1.1	124M	23.16	21.40	31.80	84.2%	45.2 ms
Torchcast 27B (Prior #1)	27B	65.10	63.20	70.40	98.1%	185.0 ms
DM-JEPA 1.5 (JEPA 2.0)	94.24M	68.92 (#1 Global)	65.69	74.55	100.0%	30.97 ms

• Tools & Automation: 0.89
• Language Understanding: 0.76
• Retrieval & Classification: 0.72
• Arts & Human Taste: 0.58
• Knowledge & Reasoning: 0.50

---

B. ISO 26262 ASIL-D Autonomous Vehicle Safety Benchmark

Evaluated across 5,000 high-risk physical autonomous vehicle scenarios under varying environmental friction (friction μ = 0.15 - 0.85) and sensor dropouts:

Metric / Safety Criterion	JEPA 1.5 Baseline	DM-JEPA 1.5 (JEPA 2.0)	Impact
Overall Decision Accuracy	83.76%	92.46% (4,623 / 5,000)	+8.70% boost
Unprotected Left Turns	53.76%	100.00% (638 / 638)	Flawless gap arbitration
Phantom Braking Rate	0.09% (4 false alarms)	0.00% (0 / 4,214 clear roads)	Zero false alarms
Missed Fatal Crash Invariants	0.00%	0.00% (0 collisions / 786)	Zero catastrophic failures
Fail-Safe Invariant Precision	99.49%	100.00%	Zero false emergency halts
Black Ice Adaptation (μ < 0.25)	90.19%	97.12% (1,079 / 1,111)	ABS straight-line enforcement
Reaction Latency (P50)	15.90 ms	15.43 ms (64.8 Hz)	Real-time automotive compliance
P99 Reaction Jitter	7.77 ms	6.25 ms	Deterministic control execution

---

4. Model Specifications

Parameter	Value	Description
Total Parameters	94,242,819 (94.24M)	Total learned weights across backbone and JEPA heads
Weights Footprint	359.6 MB / 487.5 MB	PyTorch checkpoint / Uncompressed SafeTensors on disk
Hidden Dimension (d_model)	1,024	Internal feature vector width per state (16 attention heads × 64 dimensions)
Layers / Heads	12 layers, 16 attention heads	Backbone transformer depth and parallel attention streams
Vocabulary Size	32,768	Hardware-tokenized discrete state and action vocabulary
Max Context Capacity	32,768 state / 512 option	Maximum token capacity for scene trajectory and candidate actions
Factorization Rank	p = 32	Cartesian Block Butterfly sub-quadratic projection rank

---

5. Loading the Weights

The model weights can be inspected and loaded directly via SafeTensors or PyTorch:

from safetensors.torch import load_file
import json

# 1. Load architecture configuration
with open("config.json", "r") as f:
    config = json.load(f)

# 2. Load safe open weights
weights = load_file("model.safetensors")
print(f"Loaded {len(weights)} weight tensors. Hidden dimension: {weights['backbone.embeddings.weight'].shape[1]}")

---

6. License & Citation

@misc{dangerlabs2026dmjepa,
  author = {Danger Labs},
  title = {DM-JEPA 1.5: System 1 Predictive Decision Architecture with Riemannian Metric Scoring},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/DangerLabs/dm-jepa-1.5}}
}