The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
Deterministic Worker Construction Before Inference: Measuring Repeated State Serialization and Its Task-Dependent Limits
ADI Research Series — Paper 3
Branden Laskowski · 2026 · Open-access preprint
Research question
What changes when stable worker state is assembled before inference and only the bounded task packet is supplied to the model?
Contribution
This preprint reports a frozen local synthetic benchmark comparing repeated full worker prompting with a HOPE assembled-worker path. The measured result is deliberately bounded to the named benchmark and does not establish universal model equivalence, financial or energy savings, or independent replication.
Source of record
- DOI: https://doi.org/10.5281/zenodo.23603637
- Zenodo record and public evidence companion: https://zenodo.org/records/23603637
Review boundary
The public record excludes proprietary worker packages, policy logic, raw prompts, source code, and runtime architecture. It welcomes reproductions with different models and task distributions.
- Downloads last month
- -