RELDEN Development Update — September 28, 2026

Following the previous training and evaluation work, the goal was not to start another training run as quickly as possible. Instead, the foundation for RUN003 was systematically reviewed and corrected.

The result is meaningful progress — but RUN003 is not yet cleared for training.

What changed this week

Preparation for RUN003 moved forward significantly.

RUN003 is the next controlled training step for RELDEN-8B. The broader objective remains the same: adapt an already capable language model more closely to the way RELDEN works without unnecessarily reducing its general capabilities.

The training dataset was revised and then subjected to an independent technical review.

That review initially identified several issues, including gaps in German-language coverage and requirements related to isolation and reproducibility.

Instead of accepting these issues as “good enough,” the dataset and supporting preparation were revised.

A subsequent targeted review confirmed that important issues have now been resolved. The required German scenario coverage passes the current review, and the intended dataset is correctly bound to the training configuration.

This brings RUN003 closer to being ready for execution.

A small technical issue with an important consequence

One reproducibility issue remains.

Testing in a fresh Windows environment revealed that a configuration file can be represented differently depending on how line endings are handled by the operating system and development environment.

To a person, the contents appear effectively identical. To a system performing strict file-integrity verification, however, the files are different.

That distinction matters.

RELDEN's training process is being designed so that an approved experiment can be reproduced reliably. A training configuration should not only work on the original development machine. The same approved inputs must also be recognized consistently in a clean environment.

The validation system detected this difference and correctly prevented RUN003 from being approved.

This is currently the remaining known reproducibility blocker identified by the targeted review.

Why we are not simply continuing anyway

With AI training, it is relatively easy to focus only on the resulting model and a set of benchmark numbers.

For RELDEN, that is not enough.

We want to be able to establish which data was used, which configuration produced a model, and whether two experiments were actually performed under equivalent conditions.

That is why training data and configuration are fixed and validated before a run begins.

If those checks fail, training should not simply continue regardless.

This can make individual development steps slower, but it reduces a much larger risk: comparing results from experiments whose starting conditions were not actually identical.

What previous training work has taught us

The work on RELDEN-8B has also reinforced an important lesson: specialization alone is not enough.

An adapted model can improve on specific RELDEN tasks while simultaneously losing capabilities that the original model already handled well.

Because of that, evaluation does not focus solely on whether RELDEN-specific tasks improve.

We also examine broader capabilities, structured task execution, tool-related behavior, and other areas that matter to the eventual RELDEN system.

Those lessons are directly influencing the preparation of RUN003.

The goal of the next training run is therefore not simply to achieve higher isolated scores. It is to determine whether RELDEN can become more specialized while preserving more of the underlying model's broader capabilities.

Current development status

Completed and verified:

The RUN003 dataset has been revised and reviewed again. German scenario coverage now passes the targeted review, and the intended dataset can be correctly bound to the training configuration.

The existing validation mechanisms are also successfully detecting changes to approved files rather than silently accepting them.

Not yet completed:

RUN003 has not yet been approved for the actual training run.

A platform-dependent difference affecting one configuration file currently prevents full reproducibility in a clean Windows environment.

This also means there is no completed RUN003 model and there are no RUN003 training results to report yet.

Those results will only be published after the training run has actually been completed and evaluated.

What's next

The immediate next step is clearly defined.

First, the remaining reproducibility issue must be resolved. The RUN003 preparation will then be validated again.

Only after that validation passes should the actual training run begin.

The more important phase follows afterward: evaluation.

That evaluation will determine whether RUN003 achieves a better balance between RELDEN-specific specialization and preservation of the broader capabilities of the underlying model.

Until then, RUN003 remains a training candidate in preparation — not a completed model milestone.

That distinction is intentional.

The RELDEN Development Journal is meant to document what has actually happened, not turn every planned training run, prepared feature, or promising internal idea into a completed achievement.

RELDEN remains under active development. Sometimes progress means successfully completing a new capability. At other times, it means the system correctly stopping a process before a small technical inconsistency turns into an experiment that cannot be reliably reproduced.

LST — Laubenstein Technologies

Building RELDEN — A More Capable You.