OmegaZero v8.1, recorded May 18, 2026
3-task Permuted MNIST
- Task sequence
- 3 permutations × 10 digits
- Final mean / minimum
- 98.05% / 97.56%
- Replay
- Minimal replay
Developing stable intelligence
for the next generation
At Temporal Media, we make things that work and study what it takes to keep them working as conditions change. Omega Zero Theory™ brings together our research into stability, control and recovery across systems, including AI that can keep learning while retaining what it already knows.
Our continual-learning experiments examine how well models retain earlier learning as they take on new tasks. The results below report final average task accuracy with minimal replay.
3-task Permuted MNIST v8.1
98.05%
Final average task accuracy
CIFAR-10 v14.4
92.47%
Final average task accuracy
CIFAR-100 v16.5
79.79%
Final average task accuracy
These are documented internal results from May 2026. View the research results.
Our international patent application was published on as WO/2026/185606 — view publication details.
Omega Zero Theory™
Systems have to adapt to changes in their inputs, workloads and surroundings. We study how they can do that while keeping the behavior that makes them useful.
Omega Zero Theory™ brings together our work on stability, control and recovery across systems. It guides the development of technology that can adapt to change and continue to work reliably. Our continual-learning experiments explore one part of that wider purpose.
Explore the researchResearch / Labs
Our research into stability extends across the application areas described in our PCT, including artificial intelligence, robotics and autonomous control, signal and audio processing, distributed timing networks and secure communications. The application also identifies energy management, quantum-computing environments, physiological and wearable systems, and financial and predictive modeling. Across these fields, we are interested in the same practical outcome: systems that keep working reliably as conditions change.
Continual learning is one application of that wider research. A model learns new tasks while retaining useful performance on earlier ones. We are working toward stable learning over long horizons with near-zero replay, and the results below document selected experiments along that path.
Results from our continual-learning experiments
These May 2026 results report final accuracy across each run’s task sequence. The mean is the average across those tasks, and the minimum is the lowest task score. All four runs used minimal replay.
OmegaZero v8.2, recorded May 18, 2026
5-task Permuted MNIST
- Task sequence
- 5 permutations × 10 digits
- Final mean / minimum
- 96.72% / 95.03%
- Replay
- Minimal replay
OmegaZero v14.4, recorded May 21, 2026
CIFAR-10
- Task sequence
- 5 tasks × 2 classes
- Final mean / minimum
- 92.47% / 88.65%
- Replay
- Minimal replay
OmegaZero v16.5, recorded May 23, 2026
CIFAR-100
- Task sequence
- 10 tasks × 10 classes
- Final mean / minimum
- 79.79% / 70.70%
- Replay
- Minimal replay
The earlier MNIST experiment with zero replay
In the two-task Permuted MNIST experiment recorded on May 17, 2026, OmegaZero v4.5.1 learned the second task with zero replay of earlier training examples. Its accuracy on the first task decreased by 0.61 percentage points.
About these scores
These are internal development scores from experiments that used test-set feedback. MNIST tasks classified all ten digits; CIFAR scores report accuracy within each task’s class group.
Patent & intellectual property
Our patent work accompanies the development of Omega Zero Theory™ and its approaches to calibration, control and recovery. The international application was published on September 10, 2026. Its publication details are listed here for researchers and prospective licensing partners.
Published · Patent pending
INTERNATIONAL PUBLICATION NO.
WO/2026/185606
- Publication date
- International Application No.
- PCT/IB2026/000270
- Applicant
- Temporal Media LLC
- Inventor
- Alan A. Auclair
Publication of an application does not constitute a granted patent.
Search WIPO PATENTSCOPE (opens in a new tab) using publication number WO2026185606.
Patent & licensing inquiriesAbout Temporal Media
Temporal Media LLC is an independent veteran-owned company established in Kingwood, Texas, in 2025. Founded by Alan A. Auclair, it develops and manages original intellectual property across technology, publications, media and creative work.
Our technology research centers on stability across systems. Through Omega Zero Theory™, we investigate how calibration, control and recovery can support systems that remain useful as their conditions change. Continual learning is one of the applications through which we develop and test that work.
Research and practical development belong together in this process. Ideas take shape in working systems, and the results of those systems guide what we do next.
Alan A. Auclair
Founder & inventorTalk with us about the work
For research collaborations, business questions or licensing discussions, contact Temporal Media.
Temporal Media LLC
2261 NorthPark Dr., Suite 203
Kingwood, TX 77339
2261 NorthPark Dr., Suite 203
Kingwood, TX 77339