Gyro Decision 1.0
Gyro Decision is a specialized AI decision model designed for fast, bounded decision-making inside AI systems and agents.
It operates over a context and a candidate set to perform decisions such as selection, ranking, scoring, routing, classification, and abstention.
Rather than replacing a general-purpose LLM, Gyro Decision is designed to complement one: the LLM can handle reasoning, planning, and generation, while Gyro Decision handles repeated bounded decisions.
Model Details
| Property | Value |
|---|---|
| Model | Gyro Decision 1.0 |
| Architecture | Transformer-based decision model |
| Total parameters | 450,488,198 |
| Trainable parameters | 129,300,000+ |
| Model weights | ~1.7 GB |
| Training precision | BF16 |
| Training hardware | NVIDIA A100 40GB |
| Training epochs | 2 |
| Effective batch size | 64 |
| Backbone learning rate | 1e-6 |
| Decision-module learning rate | 1e-5 |
| Context type | Candidate-set aware |
| Primary use | AI agents and decision systems |
What It Does
Gyro Decision takes a context/state together with a set of candidate options and evaluates the available alternatives.
Context / State
+
Candidate Set
↓
Gyro Decision
↓
Scores / Probabilities
↓
Select / Rank / Route / Abstain
The candidate set is part of the decision context, allowing the model to evaluate alternatives rather than generating an unconstrained response.
Supported Decision Types
Gyro Decision is designed to support bounded decision tasks including:
- Choice
- Multi-choice
- Ranking
- Scoring
- Routing
- Abstention
Intended Use
Potential applications include:
- Coding-agent file selection
- Test selection
- Change-scope decisions
- Tool selection
- Model routing
- Retrieval ranking
- Browser and computer-use action selection
- Research evidence selection
- Cybersecurity prioritization
- DevOps/SRE routing
- Agent escalation and abstention
Benchmark Results
General Decision Performance
Frozen Jev-derived evaluation
- Accuracy: 73.92%
- Top-3 accuracy: 94.25%
- MRR: 0.8369
- Evaluation size: 1,200 examples
The frozen evaluation uses the same fixed examples, candidate sets, labels, and evaluation protocol for cross-version comparison.
Engineering Decision Performance
Clean held-out engineering benchmark
- Accuracy: 81.00%
- Top-3 accuracy: 92.00%
- MRR: 0.8617
- Evaluation size: 750 examples
SWE-bench test
- Accuracy: 80.69%
- Evaluation size: 963 examples
SWE-Gym test
- Accuracy: 100.00%
- Evaluation size: 250 examples
Benchmark Comparability
The Frozen Jev-derived evaluation is a fixed evaluation set used to measure general decision performance.
Published Jev/Laya results use different evaluation setups, datasets, and/or protocols. Their figures should therefore not be interpreted as a controlled head-to-head comparison with Gyro Decision.
Training Configuration
Gyro Decision 1.0 was trained using a mixed dataset containing general decision examples and software-engineering decision examples.
Dataset
| Split | Examples |
|---|---|
| Training | 45,355 |
| Validation | 4,535 |
| Calibration | 3,969 |
| Test | 2,835 |
| Total | 56,694 |
Dataset composition:
- Approximately 65% general decision data
- Approximately 35% engineering decision data
Training Configuration
- Epochs: 2
- Total parameters: 450,488,198
- Trainable parameters: ~129.3M
- Precision: BF16
- Effective batch size: 64
- GPU: NVIDIA A100 40GB
- Backbone learning rate: 1e-6
- Decision-module learning rate: 1e-5
- TF32: Enabled
- Gradient accumulation: 8
- Per-device batch size: 8
Abstention
Gyro Decision includes a dedicated abstention mechanism.
Instead of forcing a decision when confidence is insufficient, the model can indicate that the decision should be escalated to another model, agent, or human.
For production applications, abstention and confidence thresholds should be calibrated using application-specific data.
Inference
A typical integration looks like:
AI Agent / LLM
│
│ Context + Candidate Set
▼
┌──────────────────┐
│ Gyro Decision │
└────────┬─────────┘
│
├── Selected candidate
├── Candidate scores
├── Ranking
└── Abstain
This allows a larger AI system to delegate repeated bounded decisions to a dedicated decision model.
Limitations
Gyro Decision is a specialized decision model and is not a general-purpose conversational LLM.
Performance can depend on:
- Candidate-set quality
- Number of candidates
- Decision type
- Domain
- Context quality
- Confidence thresholds
- Calibration
Benchmark results should not be interpreted as guaranteed performance for a particular application.
Application-specific validation is recommended before deployment in production, particularly for high-impact or safety-critical workflows.
Model Size
The released checkpoint contains:
450,488,198 parameters
with model weights of approximately:
1.7 GB
The model is intended to provide substantially smaller inference requirements than large generative language models while specializing in bounded decision tasks.
License
See the repository license and applicable upstream model, code, and dataset licenses before redistribution or commercial deployment.
Citation
Gyro Decision 1.0
Gyroscape
2026
About Gyroscape
Gyro Decision is part of the Gyroscape AI ecosystem, focused on building AI systems that combine reasoning, decision-making, coding, design, and autonomous workflows.
Gyro Decision — a dedicated decision layer for AI systems.